Development and implementation of an advanced

Development and implementation of an advanced auditory evoked potentials recording system. New strategies for high-rate stimulation

Joaqu´ın Tomás Valderrama Valenzuela Department of Signal Theory, Telematics and Communication University of Granada

Summary of the dissertation submitted for the degree of Doctor September 2014

Authorization ´ Dr. Angel de la Torre Vega and Dr. José Carlos Segura Luna, professors at the Department of Signal Theory, Telematics and Communications of the Superior Technical School of Computer Science and Telecommunication Engineerings of the University of Granada, as supervisors of this doctoral thesis, guarantee that the present work entitled:

DEVELOPMENT AND IMPLEMENTATION OF AN ADVANCED AUDITORY EVOKED POTENTIALS RECORDING SYSTEM. NEW STRATEGIES FOR HIGH-RATE STIMULATION

has been carried out and written by the Ph.D. student Mr. Joaqu´ın Tomás Valderrama Valenzuela, giving our authorization for its presentation in Granada, September 30th, 2014.

´ Dr. Angel de la Torre Vega

Dr. José Carlos Segura Luna

Abstract

Auditory evoked potentials (AEP) represent the nervous activity of different sections of the auditory pathway in response to a stimulus. Auditory brainstem response (ABR) and middle latency response (MLR) are AEPs used in clinics and research centers worldwide for both clinic and research purposes. In this doctoral thesis, a flexible, high-performance and cost-efficient ABR and MLR recording system is designed, implemented and evaluated. This system is appropriate for research purposes. Additionally, this thesis presents new strategies in high-rate stimulation. The randomized stimulation and averaging (RSA) technique allows the recording of ABR and MLR signals at high stimulation rates, overcoming limitations imposed by the conventional stimulation technique and presenting certain advantages in comparison with other techniques of equivalent functionality. RSA presents, however, the limitation that the jitter must be greater than the dominant period of the AEP components. The iterative-randomized stimulation and averaging (I-RSA) technique maintains the main properties of RSA, while softening its constraints significantly. The AEP recording system and the advantages of RSA have allowed for the first time in humans the study of the fast and slow mechanisms of neural adaptation, obtaining conclusions consistent with previous animal studies. Finally, this thesis presents the fitted parametric peaks (FPP) procedure, which provides an automatic quality assessment of ABR signals and parameterization of its main components in terms of amplitude, latency and width. The experimental methodology, the description of the developed techniques and procedures, the results obtained in the experiments and the significance of the findings of this work are presented in detail in this doctoral thesis.

Extended summary Auditory evoked potentials (AEP) represent the nervous activity of the auditory pathway in response to a stimulus. There are different types of AEPs according to their generator site. The auditory brainstem response (ABR) is an AEP generated in the auditory nerve and in the brainstem. This signal is composed of a number of voltage peaks which are identified with sequential Roman numerals, occurring during the first ten milliseconds from stimulus onset. Although up to seven peaks can be identified, the most prominent components are waves I, III and V. The middle latency response (MLR) is generated in the thalamocortical auditory system. MLR signals include the components Na , Pa , Nb and Pb , which appear during the first hundred milliseconds from stimulus onset. The recording of these signals is used in clinics and research centers worldwide to objectively assess the auditory system of a subject, to determine hearing threshold, to detect certain auditory pathologies, and even, to determine brain death. The AEP recording process consists in the presentation of a number of stimuli to a test subject and the recording of the electrical activity associated with each stimulus (sweep) through surface electrodes placed on the skin at different positions on the head. The low amplitude of these signals requires a large amplification. The recorded auditory evoked potentials are usually contaminated by artifacts of different nature, such as neuro-muscular activity of the subject, electronic noise derived from amplification and electromagnetic and radio-frequency interferences. The most common method to reduce the effects of these interferences is the average of a large number of auditory responses in order to improve the signal-to-noise ratio (SNR). There exist several commercial AEP recording systems, however, most of them are designed for specific audiology applications, and do not provide the flexibility required in research.

iii

The conventional stimulation technique (CONV) consists in the presentation of a number of stimuli periodically presented, i.e., with a fixed inter-stimulus interval (ISI). This technique presents the limitation that the ISI must be greater than the averaging window (i.e, 10 ms for ABRs and 100 ms for MLRs) to avoid the responses to be overlapped. Otherwise, recovering the transient AEP would be mathematically impossible. This limitation means that ABR signals cannot be obtained with the CONV technique at stimulation rates higher than 100 Hz nor MLRs at rates higher than 10 Hz. However, the recording of these signals at higher rates presents certain audiological and neurological interest, such as the analysis of neural adaptation phenomena and the detection of certain pathologies. There are several techniques that allow the recording of AEPs at high stimulation rates, in which sweeps are usually overlapped. These techniques use different deconvolution techniques, being the most important the following techniques: maximum length sequence (MLS), quasiperiodic-sequence deconvolution (QSD), continuous loop averaging deconvolution (CLAD) and least-squares deconvolution (LSD). These techniques process blocks of sweeps and assume that each stimulus evoke the same response, which is a limitation to design certain experimental procedures. The analysis of AEPs is not usually straightforward because of noise and the variability on the morphology of AEPs (due to the characteristics of the subject and the stimulation settings, such as stimulus level or rate. Response detection is typically carried out subjectively by an experienced audiologist. However, the use of automatic procedures in the assessment of these signals incorporates several advantages, such as reducing the bias derived from subjective interpretations of these signals, providing a universal systematic analysis. There are several techniques that provide an objective analysis of AEPs, being the most commonly used in clinical systems (a) the technique based on the correlation coefficient (r), which evaluates the presence of the response considering the reproducibility of two signals recorded in the same conditions; (b) the technique based on the ratio of the power of the averaged recording and the variance of a single point across sweeps (Fsp), which evaluates the quality in terms of SNR; and (c) the cross correlation of a test signal with a predefined template (Cross Corr). However, the technique based on the correlation coefficient requires a second recording,

iv

which doubles the recording test time; the technique based on Fsp requires the recording of the full electroencephalogram to be implemented offline; and the use of the Cross Corr technique is constrained to a particular protocolized stimulation and recording settings. Furthermore, the Fsp and r techniques would provide an inaccurate positive evaluation if a recording is contaminated by a strong artifact synchronized with the stimulus. The research activities carried out in this doctoral thesis belong to one out of the following four main fields: (a) the development of an AEPs recording system appropriate for research purposes; (b) the development of stimulation and processing techniques that allow AEPs to be recorded at high stimulation rates; (c) the application of the proposed AEP recording system and high-rate strategies to carry out a study of neural adaptation; and (d) the development of automatic procedures that provide an objective evaluation of the quality of AEP signals in order to, among other applications, assess the performance of the AEP recording system and high-rate stimulation techniques developed in this thesis. In this doctoral thesis, a modular architecture for recording AEP signals was conceived. This architecture is made of software and hardware independent modules, each of them performing a specific function either in the stimulation sequence generation, in the presentation of the stimuli, in the recording of the electroencephalogram (EEG) or in the digital processing of the auditory responses. The modular conception of this architecture provides the flexibility required in many research activities. Furthermore, a high-gain and low-noise amplifier, appropriate for AEPs recording, was designed, implemented, characterized and evaluated. This amplification unit was compared with the commercial amplifier portable bioamplifier BMA-200 (CWE, Inc., Ardmore, PA), presenting the prototype a notably superior performance. In this framework, an ABR and MLR recording system was developed based on the proposed modular architecture and on the amplification unit. Compared to the commercial recording system GSI Audera (Grason-Stadler Inc., Eden Prairie, MN), the developed AEP recording system: (a) presents a higher degree of flexibility, which results appropriate for research purposes; (b) is portable, which removes the need for being plugged to the electrical network; (c) the implementation cost is significantly lower; and (d) allows the recording of similar quality ABR and MLR signals. This system has allowed all

v

experimental designs of this doctoral thesis to be implemented. The development, characterization and evaluation of the performance of this AEP recording system is proposed, additionally, as a laboratory activity using the pedagogical model based on experiential learning, which may have educational advantages in certain technical profile degrees with an approach to biomedical applications, such as Electronic Engineering, Telecommunication Engineering, Computer Sciences and Audiology. This thesis presents the randomized stimulation and averaging (RSA) and iterative-randomized stimulation and averaging (I-RSA) techniques, which allow the recording of ABR and MLR signals at high stimulation rates. The RSA technique consists in the average of auditory responses corresponding to a burst of stimuli whose ISI varies randomly according to a predefined probability distribution along the entire stimulation sequence (randomized stimulation). This technique includes a digital blanking process that considers as null values any EEG samples contaminated by the stimulus artifact. In contrast to the MLS, QSD, CLAD and LSD, RSA does not perform deconvolution, thus this technique must deal with interference derived from overlapping adjacent responses. This type of interference can be reduced by averaging, provided that the jitter1 of the stimulation sequence is large enough to enable positive and negative components of this interference to be cancelled out. The I-RSA technique includes estimation of the interference associated with overlapping responses, subtraction of this interference from the EEG recorded, and re-estimation of the AEP. Each iteration leads to a better estimate of the AEP. As in RSA, the generation process of stimulation sequences in I-RSA is based on randomized stimulation. The performance of the RSA technique was evaluated through a comparison of the quality of ABR responses acquired with QSD and CONV. The results of this study suggest: (a) that the quality degrades when the ISI decreases (when the stimulation rate increases) because of the reduction of the amplitude of the response; (b) that the quality of ABR signals recorded using RSA and CONV is very similar, but with the advantage for RSA of being able to record ABR at rates higher than 100 Hz; and (c) that the quality of the responses recorded with RSA is slightly better 1

The jitter of a stimulation sequence measures the amount of dispersion of the ISI in contrast to a periodical presentation of stimuli.

vi

than that of the QSD responses, especially at higher stimulation rates. The performance of the I-RSA technique was analyzed by two experiments with both real and computer simulated ABR and MLR signals obtained at different rates and jittering conditions. The results of this study indicate that RSA can be efficiently used for recording AEPs provided that the amount of jitter of the randomized stimulation sequence is greater than the dominant period of the ABR/MLR components. When this premise is not fulfilled, positive and negative components of the interference associated with overlapping responses cannot be cancelled out in the averaging process, and the resulting AEP will not be reliable. The performance of I-RSA maintains all the advantages of RSA: (a) it allows the jitter distribution to be controlled with precision, (b) stimulation sequences are easy to generate, and (c) it allows responses to be processed separately. Additionally, I-RSA significantly smooths the restrictions of a minimum jitter. The AEP recording system and the high-rate stimulation techniques developed in this thesis were used to carry out a study that has allowed for the first time in humans the study of the fast and slow mechanisms of neural adaptation. In this study, ABR signals were recorded as reference by the RSA technique at a low stimulation rate (long ISIs) and at a high stimulation rate (short ISIs). In addition, stimulation sequences composed of both long and short ISIs randomly distributed were used to obtain the ABR test signals. The separated responses technique developed in this study allowed the categorization of sweeps according to their previous ISI. Therefore, two separated ABR signals could be obtained by the average of the sweeps belonging to the long ISI and short ISI groups. Two hypotheses were taken under consideration in this study. If the fast mechanisms of adaptation prevailed over the slow mechanisms, the separated ABR signals would be similar to their corresponding reference ABR signals (both separated ABRs would be different) since the morphology of the response would be influenced in a greater extent by the ISI of the preceding stimulus. On the other hand, if the slow mechanisms of adaptation prevailed over the fast mechanisms, the separated ABR signals would be different to their corresponding recorded ABR signals and both separated ABRs would be similar because the morphology of the response would be mainly determined by the averaged stimulation rate of several preceding stimuli (but not by the ISI of the preceding stimulus). The results of this study

vii

indicate that most of the subjects present a situation in between both hypotheses, which suggests that both fast and slow mechanisms of adaptation influence the morphology of the auditory response. These results are consistent with previous animal studies from different researchers. The performance of the AEP recording system and the high-rate stimulation strategies developed in this thesis can be assessed by the analysis of AEP signals. However, the subjective analysis of these signals manifests problems of standardization and the available automatic techniques present certain limitations. In this framework, this thesis describes and evaluates the performance of the fitted parametric peaks (FPP) technique, a new approach for automatic quality assessment and peak parameterization based on the use of templates. The approach of FPP consists of the search of the latency, width, and amplitude of a parametric peak, similar in morphology to an ABR wave that best fits the most robust waves of the ABR, waves III and V. The performance of FPP was evaluated in this study by two experiments. In the first experiment, the latencies and amplitudes of waves III and V were estimated manually by an audiologist and automatically by FPP in ABR signals obtained from eight normal hearing subjects at different stimulation rates. This analysis shows that the FPP technique provides more consistent results than the manual procedure, possibly due to the fact that FPP bases the estimation of the parameters considering an interval of the response, rather than isolated samples, which makes the FPP technique less sensitive to noise. In the second experiment, the performance of FPP was contrasted with the most common automatic quality evaluation procedures: the correlation coefficient (r), the Fsp, and the cross correlation with a predefined template waveform (Cross Corr). These automatic quality evaluation techniques were compared to a subjective evaluation provided by five experts. The results of this test revealed that although all automatic techniques present high correlation coefficients with the averaged subjective assessment, FPP remains as the technique that best approaches an averaged subjective evaluation. Furthermore, these results point out that there is an important bias among the evaluators, which evidences that the reproducibility of visual judgments is not high, revealing the convenience of using automatic procedures.

viii

Contributions of this thesis This thesis presents the following contributions: A modular auditory evoked potentials recording architecture. The modular

nature of this architecture allows optimal recording settings to be implemented for specific experimental procedures, which is interesting for research purposes. Furthermore, this thesis contributes with (a) the design of a high-gain and low-noise amplifier, whose performance is proven to be appropriate for the recording of auditory evoked potentials; and (b) the configuration of an ABR and MLR recording system, set up by a laptop, an external sound-card, surface electrodes, insert earphones, an amplifier, a graphical software platform that allows the intuitive AEP recording process, and software methods that perform the necessary signal processing to record the AEP. The randomized stimulation and averaging (RSA) technique, which allows

the recording of auditory evoked potentials at high stimulation rates. The RSA technique consists in the average of auditory responses corresponding to a number of stimuli whose ISIs are randomly distributed according to a predefined probability distribution and the application of a digital blanking process. Digital blanking prevents EEG samples contaminated by the stimulus artifact to be processed. RSA presents a number of advantages in comparison with the conventional technique and other high-rate stimulation techniques based in deconvolution. However, RSA presents the constraint of a minimum jitter, which must be greater than the dominant period of the AEP to allow positive and negative components of the interferences associated with overlapping responses to be cancelled by averaging. The iterative-randomized stimulation and averaging (I-RSA) technique per-

forms the estimation and suppression of the interference associated with overlapping responses through an iterative process in the time domain. This technique maintains the main advantages of RSA and smooths the restriction of a minimum jitter. This allows a more flexible use of randomized stimulation.

ix

The RSA with separated responses technique allows the categorization of

auditory responses according to a predefined criterion, e.g., the preceding ISI. This technique allows the assumption that different stimuli evoke responses of different morphology. This approach allows the design of experimental procedures that cannot be addressed by other techniques based in deconvolution. A procedure to study the fast and slow mechanisms of adaptation through a

non-invasive technique, appropriate for its use in humans. This procedure is based on the comparison of the morphology (amplitudes and latencies of the main components) of (a) reference ABR signals, obtained directly from stimulation sequences made up of stimuli with long and short ISIs, and (b) test ABR signals, obtained with the RSA with separated responses technique from stimulation sequences with long and short ISIs randomly distributed. This procedure was used in a study, which has revealed that both fast and slow mechanisms of adaptation are present in humans. The fitted parametric peaks (FPP) procedure allows the quality and param-

eterization of the main components of ABR signals in terms of amplitude, latency and width to be automatically evaluated. This procedure consists in the search of the amplitude, latency and width of an artificially synthesized peak similar to the morphology of an AEP component. FPP is implemented with a computationally efficient algorithm, reducing the fitting problem to a one dimensional search.

x

Conclusions With regard to the recording system, the main conclusions of this doctoral thesis are presented below: The proposed auditory evoked potentials recording architecture using in-

dependent modules provides a high-grade of flexibility. This is appropriate for research purposes, and additionally, it allows the comparison of the performance of different implementations for a given module. The amplifier unit developed in this thesis presents an adequate perfor-

mance in the recording of auditory evoked potentials, notably superior to the commercial amplifier BMA-200 (CWE, Inc., Ardmore, PA). The ABR and MLR recording system presents an adequate performance,

which is comparable with the commercial clinical system GSI Audera (GrasonStadler Inc., Eden Prairie, MN), and which can be implemented with lowcost elements, providing a high grade of flexibility. This system is appropriate for both research and education purposes. With regard to the high-rate stimulation techniques: The RSA technique presents the following advantages: compared to the

conventional technique, RSA allows the recording of AEPs at high stimulation rates, in which auditory evoked responses are, generally, overlapped; in comparison with other techniques based on deconvolution, the main advantages of RSA are: (a) it allows a precise control of the jitter, (b) stimulation sequences are easy to generate, since they are not constrained to restrictions in the frequency domain, and (c) it allows an individual processing of auditory responses. The major limitations of the RSA technique are: (a) the jitter of the stimulation sequence must be greater than the dominant period of the recorded AEP, and (b) the duration of digital blanking and the amount of jitter of the stimulation sequence must allow the average of a predefined minimum number of the available samples across the averaging window.

xi

In RSA, when the jitter of the stimulation sequence is lower than the period

of the recorded AEP, constructive or destructive interference contaminate the recording, and these effects cannot be reduced by averaging. On the other hand, when the duration of digital blanking and the amount of jitter do not allow averaging at least the predefined minimum number of the available samples, significant differences in quality may arise between different segments of the response. Compared to QSD, the RSA technique presents a better performance, pos-

sibly associated to two mechanisms (a) the quality of the recordings in QSD is strongly influenced by the selection of the stimulation sequence (qsequence), and (b) RSA performs a more efficient implementation of artifact rejection techniques, since this technique processes individual responses, instead of blocks of responses. I-RSA maintains all the advantages of RSA, while eliminating the need of

digital blanking and smoothing the restriction of a minimum amount of jitter in the stimulation sequences. The RSA technique allows a significant reduction of recording test time at

high intensity levels. However, at low intensity levels, in which generally the identification of the components of ABR signals is not direct, high stimulation rates reduce the amplitude of these components, difficulting response detection even more. Thus, the benefits of high stimulation rates to reduce recording time are still not clear in applications such as hearing screening or threshold estimation. With regard to the study of neural adaptation: The RSA with separated responses technique allows the analysis of the fast

and slow mechanisms of adaptation in humans through the analysis of the morphology of ABR signals obtained at fast stimulation rates. The analysis of this study shows evidences of both the fast mechanisms of

adaptation, with an associated time constant of a few milliseconds, and of the slow mechanisms of adaptation, whose time constant is greater than

xii

20 ms. These evidences are consistent with other previous animal studies carried out by invasive techniques. With regard to the automatic quality evaluation technique: The FPP technique allows a computer efficient implementation, providing

an automatic evaluation of the quality and peaks parameterization in real time. The estimation of the amplitude and latency provided automatically by

the FPP technique is comparable with the estimation of these parameters carried out manually by an experienced audiologist. Compared to the automatic techniques based on the correlation coefficient

(r), Fsp, and cross correlation with a predefined ABR template (Cross Corr), FPP remains as the technique that best approximates the averaged subjective quality evaluation provided by five experienced audiologists. There exists a significant bias between the subjective evaluations of the

quality provided by five experienced audiologists, which supports the use of automatic procedures. The FPP technique presents several advantages in research applications:

(a) it may substitute the manual labeling of the components of AEPs in clinical reports, (b) it may help provide an automatic interpretation of ABR signals, (c) it may improve the automatic process of stopping averaging when an ABR of sufficient quality is recorded, making a more efficient use of the recording time, and (d) it may help perform objective comparisons of the performance of different stimulation techniques or artifact rejection techniques.

xiii

Scientific production Peer-review publications in JCR journals 1. Valderrama, J., Alvarez, I., de la Torre, A., Segura, J.C., Sainz, M., Vargas, J.L. Educational approach of a BAER recording system based on experiential learning. Technics Technologies Education Management (2011), vol. 6, no. 4, pp. 876-889. 2. Valderrama, J., Alvarez, I., de la Torre, A., Segura, J.C., Sainz, M., Vargas, J.L. Recording of auditory brainstem response at high stimulation rates using randomized stimulation and averaging. Journal of the Acoustical Society of America (2012), vol. 132, no. 6, pp. 3856-3865. 3. Valderrama, J., de la Torre, A., Alvarez, I., Segura, J.C., Thornton, A.R.D., Sainz, M., Vargas, J.L. A study of adaptation mechanisms based on ABR recorded at high stimulation rate. Clinical Neurophysiology (2014), vol. 125, no. 4, pp. 805-813. 4. Valderrama, J., de la Torre, A., Alvarez, I., Segura, J.C., Thornton, A.R.D., Sainz, M., Vargas, J.L. Automatic quality assessment and peak identification of auditory brainstem responses with fitted parametric peaks. Computer Methods and Programs in Biomedicine (2014), vol. 114, no. 3, pp. 262-275. 5. Valderrama, J., de la Torre, A., Alvarez, I., Segura, J.C, Sainz, M., Vargas, J.L. A flexible and inexpensive high-performance auditory evoked response recording system appropriate for research purposes. In press, Biomedical Engineering/Biomedizinische Technik (2014), DOI: 10.1515/bmt2014-0034. 6. Valderrama, J., de la Torre, A., Alvarez, I., Segura, J.C., Thornton, A.R.D, Sainz, M., Vargas, J.L. Auditory brainstem and middle latency responses recorded at fast rates with randomized stimulation. Under review, submitted to the Journal of the Acoustical Society of America.

xiv

Peer-review contributions to international conferences 1. Valderrama, J., Alvarez, I., de la Torre, A., Segura, J.C., Sainz, M., Vargas, J.L. Reducing recording time of brainstem auditory evoked responses by the use of randomized stimulation. Poster presentation at the Newborn Hearing Screening (NHS) congress, Cernobbio (Como Lake), Italy (June 5-7, 2012). 2. Valderrama, J., Alvarez, I., de la Torre, A., Segura, J.C., Sainz, M., Vargas, J.L. A preliminary study of the short-term and long-term neural adaptation of the auditory brainstem response by the use of randomized stimulation. Poster presentation at the Adult Hearing Screening (AHS) congress, Cernobbio (Como Lake), Italy (June 7-9, 2012). 3. Valderrama, J., Alvarez, I., de la Torre, A., Segura, J.C., Sainz, M., Vargas, J.L. A portable, modular, and low cost auditory brainstem response recording system including an algorithm for automatic identification of responses suitable for hearing screening. Oral presentation at the IEEE EMBS Point-of-Care Healthcare Technologies (PHT) conference, Bangalore, India (January 16-18, 2013), art. 6461314, pp. 180-183. 4. Valderrama, J., Morales, J.M., Alvarez, I., de la Torre, A., Segura, J.C., Sainz, M., Vargas, J.L. Automatic quality assessment and response detection of auditory evoked potentials based on response tracking. Oral presentation at the International Evoked Response Audiometry Study Group (IERASG) meeting, New Orleans, LA (June 9-13, 2013). 5. Valderrama, J., de la Torre, A., Alvarez, I., Segura, J.C., Thornton, A.R.D., Sainz, M., Vargas, J.L. Auditory middle latency responses recorded at high stimulation rates using randomized stimulation and averaging. Oral presentation at the International Evoked Response Audiometry Study Group (IERASG) meeting, New Orleans, LA (June 9-13, 2013). 6. Valderrama, J., de la Torre, A., Alvarez, I., Segura, J.C., Sainz, M., Vargas, J.L. Deconvolution of overlapping responses and frequency domain-

xv

based artifact rejection methods using randomized stimulation. Oral presentation at the International Evoked Response Audiometry Study Group (IERASG) meeting, New Orleans, LA (June 9-13, 2013). 7. Valderrama, J., de la Torre, A., Alvarez, I., Segura, J.C., Kaf, W., Sainz, M., Vargas, J.L. A more efficient use of the recording time with randomized stimulation and averaging (RSA) in hearing screening applications. Invited speaker at the 9th International Conference of the Saudi Society of Otorhinolaryngology - Head and Neck Surgery, Riyadh, Kingdom of Saudi Arabia (March 4-6, 2014).

Peer-review contributions to Spanish conferences 1. Alvarez, I., Valderrama, J., de la Torre, A., Segura, J.C., Sainz, M., Vargas, J.L. Reducción del tiempo de exploración de potenciales evocados auditivos del tronco cerebral mediante estimulación aleatorizada. Oral presentation at the XXV Unión Cient´ıfica Internacional de Radio (URSI) national symposium, Bilbao, Spain (September 15-17, 2010). 2. Valderrama, J., Franco, M., Alvarez, I., de la Torre, A., Segura, J.C. Registro de potenciales evocados auditivos mediante una arquitectura modular apropiada para propósitos de investigación. Oral presentation at the 6º Simposio CEA Bioingenier´ıa 2014, Interfaces Mente-computador y Neurotecnolog´ıas, Granada, Spain (June 12-13, 2014). Cognitive Area Networks (2014), ISSN: 2341-4243, vol. 1, no. 1, pp. 67-73. 3. Morales, J.M., Valderrama, J., Alvarez, I., de la Torre, A., Segura, J.C. Método automático de seguimiento de respuestas evocadas auditivas basado en la parametrización de series de registros. Oral presentation at the 6º Simposio CEA Bioingenier´ıa 2014, Interfaces Mente-computador y Neurotecnolog´ıas, Granada, Spain (June 12-13, 2014). Cognitive Area Networks (2014), ISSN: 2341-4243, vol. 1, no. 1, pp. 75-80.

xvi

technics technologies education management

Educational approach of a BAER recording system based on experiential learning Joaquin T. Valderrama1, Isaac Alvarez1, Angel de la Torre1, Jose Carlos Segura1, Manuel Sainz2, Jose Luis Vargas3 1

2

3

Department of Signal Theory, Networking and Communications, CITIC-UGR, University of Granada, Spain, Department of Surgery and its Specialties, University of Granada, Spain. ENT Service, San Cecilio University Hospital, Granada, Spain, ENT Service, San Cecilio University Hospital, Granada, Spain.

Abstract Brainstem Auditory Evoked Response (BAER) represents the electrical activity of the auditory nerve associated to a stimulus. The recording process of this signal is a challenging milestone that requires the development of multidisciplinary competences. Current medical devices able to record BAER are expensive and are not usually flexible enough for education purposes. This paper presents a full description of a BAER recording system designed and developed in our research group. The performance of this system is evaluated by (a) the recording of a very low amplitude synthesized signal similar in morphology to an evoked potential, (b) an analysis of the quality of the recordings in terms of the number of averaged responses and (c) the recording of BAER from eight normal hearing adults and an analysis of the influence of intensity of stimulation over the biological responses. In addition, this paper proposes the described BAER recording system to be used as a teaching tool in the context of electronics engineering, telecommunication engineering, computer sciences, and audiology studies. The lowcost and open nature of this system is appropriate for education purposes. The BAER recording system described in this paper has been used in a laboratory session. The results obtained on a satisfaction survey given to the students reveal that the use of experiential learning as pedagogical model has a great impact on the motivation of students and proves its efficiency to enhance retention and 876

improve cognitive skills. The educational advantages of our system in comparison to other biomedical devices are discussed on this paper. Key words: Brainstem Auditory Evoked Response (BAER), evoked potentials, signal processing, laboratory, MATLAB, experiential learning. 1 Introduction In the hearing process, the transmission of information from the inner ear to the primary cortical area involves the discharge of action potentials along the auditory pathway. The electrical response of the auditory system associated to a short stimulus show several waves that occur within the first 10 milliseconds after the presentation of a stimulus. These waves are known as brainstem auditory evoked potentials and correspond to synchronous discharges of action potentials from groups of neurons placed at different points at the auditory pathway [1]. Evoked potentials are identified by roman numerals as proposed Jewett in [2]. The most extended application of Brainstem Auditory Evoked Response (BAER) signals is the estimation of hearing threshold due to the non-invasive nature of the process of recording and the independence from the state of attention of the subject [1] [3]. Brainstem Auditory Evoked Potentials are recorded through the stimulation of the auditory system by presenting acoustic stimuli and the recording of the associated electrical response. This biological response is captured by surface electrodes Volume 6 / Number 4 / 2011


situated on the skin at different positions on the head. A high amplification of the recording needs to be performed due to the low amplitude of the waves (usually less than 1 µV). This signal is contaminated by several sources of artifacts such us neuro-muscular activity of the subject, noise associated to the amplifier and electromagnetic and radiofrequency interferences [4]. The methodology used to reduce the effects of these artifacts is the average of a large number of biological responses in order to improve the signal to noise ratio. Currently, there exist several clinical systems able to record BAER. However, most of them are not flexible enough for education purposes since they only allow users to control a few parameter settings, the use of artifact reduction techniques is limited, and give no access to raw recorded data [5] [6]. This paper presents in detail an inexpensive open BAER recording system. This system provides a flexible control of the parameters: users are able to specify the intensity of stimulation, set the analog to digital sample frequency, define the stimulation technique, select the number of biological responses involved in averaging, set the band-pass filter settings, or implement advanced artifact rejection techniques. In addition, this system gives total access to raw recorded data, which means that advanced processing of data can be implemented offline. Digital signal processing has been developed in MATLAB (The Mathworks, Inc.). This programming framework has been proved to be efficient as a didactic tool [7]. Three experiments have been performed to test the efficiency of the system. First, a very low amplitude digitally synthesized signal similar in morphology to an evoked potential has been recorded. Second, an analysis of the quality of evoked potentials obtained by averaging a different number of responses has been carried out. Third, BAER from eight normal hearing subjects have been recorded at different intensities of stimulation and an analysis of the latencies of the most important waves is performed. On the field of education, learning methodologies based on experiential learning can be employed as a complement of traditional textbookbased learning. This paper proposes the implementation and evaluation of a BAER recording system as a teaching tool in a laboratory frameVolume 6 / Number 4 / 2011

work. The aim of this laboratory is to help students understand specific concepts of audiology, analyze the agents that influence evoked potentials, gain experience on electronics instrumentation, improve analog and digital signal processing skills, understand the effect of impedance and noise in real life applications and learn adequate noise reduction processing methods. The performance of the laboratory activities that compose the BAER recording system is supported by the experiential learning strategy [8]. A laboratory session has held at the University of Granada in which the BAER recording system described on this paper was used to help students understand the most important concepts and theoretical models studied in audiology lectures. The laboratory comprised the performance of several activities: (a) an introduction in which the elements that compose the BAER recording system were presented, (b) the placement of the electrodes on the head, (c) the recording of BAER at different number of averaged responses and (d) the recording of BAER at different intensity levels. A satisfaction survey of the laboratory was given by the students. The rest of the paper is structured as follows. Section 2 describes in detail the hardware elements and software modules that compose the BAER recording system. Section 3 presents results of the evaluation of the system, confirming an adequate performance. Section 4 points out the skills and competences expected to be acquired by the students through the implementation of the activities of this laboratory and describes the laboratory session performed at the University of Granada. Finally, the main conclusions of this work are exposed on section 5. 2 System description This section presents the description of the system. The process of BAER recording and the elements involved in this process are introduced on section 2.1. Section 2.2 describes in detail the hardware items of this system. The software modules that compose the digital signal processing of this system are presented on section 2.3. Finally, section 2.4 performs a cost analysis of the necessary materials to deploy this system. 877


2.1 Recording of BAER The system that we have developed uses a laptop including MATLAB, an external audio sound card, an amplifier, a pair of headphones and computer routines that perform the digital signal processing. These elements are shown in figure 1. The process of BAER recording is described on the following steps: 1. A sequence of short bursts (see section 2.3.1) is generated by the laptop for stimulation and synchronization purposes. This signal is sent synchronously through the left and right outputs of an external Analog-Digital / Digital-Analog (AD/DA) sound card. The right output of the AD/DA sound card is connected to its left input, so that the system can determine the exact moment in which a stimuli is produced by the recording of the synchronization signal. The left output of the sound card is connected to a pair of headphones (section 2.2.4), through which the stimulation signal elicits the stimulation of the auditory system of a subject. 2. The electrical response is collected by three Ag/ AgCl surface disc electrodes (see section 2.2.2). The electroencephalogram (EEG) captured by the electrodes is pre-amplified by a factor G1 = 25, band-pass filtered (150 - 3000 Hz), and amplified by a factor G2 = 130. Therefore, the gain of the amplifier for the band-pass frequencies is set at about Gamp = 3250 (70 dB) (see section 2.2.1). 3. The filtered and amplified biological response and the synchronization signal are synchronously recorded by both right and left inputs of the external AD/DA sound card. Both signals are sampled at a frequency of 25 kHz and stored using 24 bits of quantization (see section 2.2.3). 4. The final step is digital signal processing. This stage includes the scaling of the input signal (section 2.3.3); the synchronization of the biological response with its corresponding stimulus; the average of auditory responses (see section 2.3.5) and the use of other additional features to enhance the SNR ratio of the evoked potentials or reach other advanced functionalities (see section 2.3.4). 5. Users have total access to raw data and a flexible control of the parameters following a few steps 878

on the interactive multimedia interface (see section 2.3.6).

Figure 1. General scheme of the system. 2.2 Hardware system This section describes the hardware elements involved in the process of BAER acquisition. These components are amplifier, electrodes, AD/ DA sound card and headphones. 2.2.1 Amplifier The low amplitude of BAER signals imposes an amplification of about 70 dB in order to be recorded by the AD/DA sound card. The amplification of this signal is performed through a biopotential amplifier that is described on this section. Figure 2 shows the electronic circuit diagram of the amplifier. The amplifier comprises four stages: preamplification, band-pass filtering, amplification and active ground circuitry. Preamplification provides a moderate gain in order to adapt the input signal and prevent saturation on later stages. Bandpass filtering stage is designed to eliminate those frequencies out of the scope of the evoked potentials, amplifying only the band of interest (1503000 Hz). Amplification is performed when the input signal is adapted and filtered, reaching the desired level of amplitude. The active ground circuit increases the Common Mode Rejection Ratio (CMRR). Electric networks generate an electric field that can be induced on the amplifier, on cables and on the subject, producing a common mode voltage in the subject. In case the subject is connected to ground, the common mode voltage would be the multiplication of the impedance of the electrode and the induced current, which Volume 6 / Number 4 / 2011


Figure 2. Electronic circuit diagram of the amplifier would flow to ground only through the subject because of the high input impedance of the amplifier. Common mode voltage is amplified, inverted and inserted back into the subject by the active ground circuit, performing an important reduction on the common mode voltage. Preamplification and amplification stages are based on the instrumental amplifier INA128 (Texas Instrument). This differential amplifier has been chosen because of its high CMRR (120 dB), low power, low noise voltage ( 8 nV / H z ), an easy control of the gain, and a lineal behavior of the gain and CMRR on the band of interest. Band-pass filtering stage is composed of four second order Sallen-Key filters (2 x high-pass & 2 x low-pass). Operational amplifiers used on band-pass filtering and active ground stages are OPA227 (Burr-Brown from Texas Instrument). OPA227 is characterized of a very low noise voltage ( 3 nV / H z ), high CMRR (130 dB) and high precision. The Bode diagram on figure 3 presents the frequency response of the amplifier. Linearity of the amplifier has also been studied. A sinusoidal input signal was inserted on the amplifier with the intention of obtaining a slightly saturated output signal. The frequency of the input signal (1.167 kHz) was chosen in order to avoid delay on the output signal. 5 milliseconds of both input and output signals were recorded. Figure 4 displays the X-Y graph of the experiment. Linearity of the amplifier can be analyzed on this figure. This figure confirms that the behavior of the amplifier is especially lineal when the input signal is bounded within the range [-500 +500] µV, a common situation on this application considering that the input signal when recording BAER do not usually exceed 10 µV. Volume 6 / Number 4 / 2011

Figure 3. Bode diagram of the amplifier

Figure 4. Linearity study. X-Y graph 2.2.2 Electrodes Auditory brainstem potentials are recorded using surface electrodes placed at different positions on the head. Electrodes perform the function of transforming ionic currents (mechanism of con879


duction of bioelectrical signals on tissues) into electrical currents. Electrodes commonly used on the field of BAER acquisition are silver - silver chloride (Ag/AgCl), composed by a silver conductor (electrode) immersed into a silver chloride salt dissolution (electrolyte). When the electrolyte is rich on chloride ions, its impedance is near to linearity. Electrolyte is used, therefore, as a mean of union between the electrode and the skin. Electrolytic paste is used between the skin and the electrode in order to reduce impedance. Since electrodes are the first elements on signal acquisition, noise generated at them is of high importance. Contact impedance between electrodes and skin must be minimized in order to reduce the noise affecting the recorded signal. It can be done by the elimination of layers of grease and dead cells. These layers can be removed by a softly scrape of the skin. Positioning electrodes on the head plays an important role. Figure 5 shows the correct position. Active electrode is placed in vertex, in middle line near hairline. Reference electrode is set on the contralateral mastoid, due to its low innervation and poor muscular tissue. Active and reference electrodes are connected to the differential input of the amplifier. Ground electrode is placed on the forehead, in middle line. This electrode acts reducing the common mode voltage gain.

reaching a very high resolution level; nevertheless, the use of 16 bits of quantization is also permitted in case it is desired to reduce the size of the file for its later processing and storing. Table 1 presents the main features and parameters of the AD/DA sound card set for this application. Table 1. Features of the AD/DA sound card. Feature Sampling rate Input range Output range Bits of quantization Input resolution

Value 25 kHz -3V / 3V -2.5V / 2.5V 24 1.36 V

2.2.4 Headphones Headphones act as an electro-acoustical transducer, transforming the electrical energy of the stimulation signal into acoustical energy (sound) that stimulates the auditory system of the explored subject. Monaural stimulation of the auditory system of a subject is provided. Circumaural standard headphones have been chosen for this application in order to reduce the effect of the background noise in the recording room. 2.3 Software system This section describes the digital processing steps necessary for the recording of brainstem evoked potentials. Figure 6 shows a diagram of the software modules involved in this process.

Figure 5. Position of electrodes on the head. 2.2.3 AD/DA sound card An external sound card has been chosen to act as interface between the digital and analog systems. This type of devices has the advantage of simplicity and a better performance than most of sound cards integrated on laptops. The external sound card is connected to the laptop through the USB port. Sampling rate has been set on 25 kHz in order to prevent aliasing. 24 bits of quantization allow 880

Figure 6. Software modules diagram. Volume 6 / Number 4 / 2011


2.3.1 Stimulation & synchronization signal generation The aim of the Stimulation & synchronization signal generation module is the generation of the signal responsible for (a) the stimulation of the auditory system of a subject and (b) the synchronization process. The stimulation & synchronization signal is composed of a train of short bursts each one consisting on a biphasic click. Biphasic clicks provoke a synchronous discharge of a large number of hair cells due to its short duration (200 µs) and its wide frequency spectrum. This fact makes easier the identification of the biological response. Parameters like stimulation rate, intensity and number of stimulus considered for averaging can be controlled by the user. 2.3.2 Stimulation & Acquisition The Stimulation & Acquisition module performs synchronously two actions: (a) the reproduction of the stimulation & synchronization signal through the left and right outputs of the AD/ DA sound card; (b) the recording of the synchronization signal and the electroencephalogram (EEG) by the left and right input of the AD/DA sound card respectively. By default, sampling frequency is set to 25 kHz and the system uses 24 bits of quantization; nevertheless, users can change these parameters in order to control the precision of the recordings. 2.3.3 Scaling A scaling process is required to measure the amplitude in volts of the brainstem auditory evoked potentials. Figure 7 shows a scheme of the elements involved in the process of scaling. AR represents the amplitude in volts of the raw electroencephalogram recorded by electrodes. GA is the gain of the amplifier. Therefore, the amplitude in volts of the signal after amplification (AV) is AV = AR*GA. AX represents the non-dimensional signal after sampling. The gain of the sound card in the process of conversion from analog to digital is represented by GS. This gain is related to the level Volume 6 / Number 4 / 2011

of intensity of the input line in the audio settings of the laptop. Medium intensity level is recommended. The function of the module Scaling is to convert the recorded signal AX into its corresponding real value in microvolts. This procedure is performed according to equation 1.

Vcalibrated ( µV ) = AX *

1 1 * * 10 GS G A

6

......... (1)

The process of scaling requires the system to be calibrated. The calibration of the system consists of determining the value of GS and GA. The gain of the amplifier for the band-pass frequencies (GA) is obtained directly from the Bode diagram. The gain of the AD/DA sound card (GS) can be measured by the recording of a few waves from a sinusoidal signal whose maximum value in volts is known (Vhi) and the correlation with its corresponding recorded value (Xhi), as the following formula indicates: GS = Xhi / Vhi.

Figure 7. Parameters involved in the scaling and calibration process 2.3.4 Advanced digital processing The process of recording BAER involves the “Advanced digital processing” step. This module incorporates optional high-level functions to enhance the SNR of the evoked potentials, such as digital filtering, the implementation of artifact rejection techniques or any other operation over the EEG. 2.3.5 Synchronization, Average, Display and Storage The module “Synchronization” uses the recording of the synchronization signal to determine the samples on the EEG where stimulation occurs. The “Average” module computes the mean of the auditory responses. The function of “Display” is to 881


visualize the result in a suitable scale. The “Storage” module saves the result and other important parameters into a file. 2.3.6 User interface The process of BAER recording can be managed from a Graphical User Interface (GUI). This interactive framework allows a non-experienced user to record evoked potentials following systematically a few simple steps. Programmers can design the structure of this multimedia platform and include any parameter that involves the process of BAER recording using the MATLAB Graphical User Interface Development Environment (GUIDE). This tool simplifies the process of designing GUIs. Figure 8 presents an example of GUI. On this figure, the first step is the generation of the stimulation & synchronization signal. Parameters like the number of stimulus, the stimulation rate and intensity level of stimulation can be controlled. Relevant information like the period of the stimulation signal and the duration of the test are displayed on the screen. On this platform, users can also visualize the evoked potentials and choose the name of the file where results are stored.

signal processing laboratory. The price list considered as reference to build this table was taken from a well-known international electronics supplier. This table points out that the cost of the system for BAER recording described on this section is much lower than other standard medical devices, whose price can reach several thousands of dollars. Table 2. Rough cost of the hardware elements of the BAER recording system. Amplifier electronics Reusable electrodes and electrolytic paste Circumaural headphones External AD/DA sound card TOTAL

Rough cost 200 $ 200 $ 30 $ 50 $ 480 $

3. Assessment of the system Three experiments have been performed in order to check the viability of the system. First, a synthesized signal similar in shape and amplitude to a biopotential has been recorded. Second, the minimum number of averaged responses in order to obtain recordings with enough Signal to Noise Ratio (SNR) to identify the main waves of BAER is analyzed. Third, brainstem evoked potentials have been recorded from eight normal hearing adults and biological conclusions have been obtained through an analysis of the latencies of the main waves of BAER. Recording of a pseudopotencial

Figure 8. An example of front-end of the system 2.4 Cost analysis Table 2 shows a rough cost of the necessary materials to set up the biometric signal processing laboratory proposed on this paper. The laptop and the license of MATLAB are not included in this table since they are common elements of any 882

Figure 9. Elements involved in the process of recording a pseudopotential The aim of this experiment is the simulation of the recording of a brainstem auditory evoked poVolume 6 / Number 4 / 2011


tential. Figure 9 schematizes the connections of the necessary elements to perform this experiment. A signal of a similar morphology to a BAER has been digitally synthesized. Figure 10 shows the “pseudopotential” that we have designed for this purpose. A high stimulation artifact is introduced on the pseudopotential to test a possible saturation of the amplifier. A burst of pseudopotentials (VA) and a synchronization signal (VS) are synchronously sent through both left and right outputs of the AD/DA sound card (Figures 10 and 11). The amplitude of the pseudopotential has been reduced by a voltage divider of a factor 105. Zout represents the output impedance of the source since the input impedance of the amplifier is considerably higher (1 MΩ). The output impedance simulates the contact impedance of the electrodes with the skin. The noise of the input signal of the amplifier increases along with the output impedance Zout. The burst of low-amplitude pseudopotentials is amplified, recorded by the AD/DA sound card, and digitally processed following the same procedure as obtaining BAER (section 2.3).

Figure 10. Synthesized pseudopotential On this experiment, the amplitude of the wave V on the burst of pseudopotentials in VA was 50 mV in order to have 0.5 µV after the voltage divider. The output impedance was set on Zout = 4.7 kΩ, simulating a common contact impedance between electrodes and skin. The recorded pseudopotential is shown in figure 12. Digital blanking was used to eliminate the effect of the stimulation artifact. No saturation of the amplifier was observed. Although there exist some differences in morphology in comparison to the synthesized pseudopotential Volume 6 / Number 4 / 2011

(figure 10), every wave on the recorded pseudopotential have been retrieved, remain on the same latency and are easily identified. This fact entails that the system is able to record signals of amplitudes lower than 1 µV.

Figure 11. The burst of pseudopotentials (VA) and the synchronization signal (VS) are sent synchronously through the left and right outputs of the AD/DA sound card

Figure 12. Retrieved pseudopotential SNR analysis Suppression of non-synchronous artifact is based on the average of biological auditory responses. A large number of responses increases the signal to noise ratio, but extends the necessary time for recording. Therefore, we have performed an analysis to find the appropriate number of averages that leads to high quality recordings spending a reasonable exploration time. 883


Figure 13 presents recordings of BAER obtained from a normal hearing adult at an intensity of 70 dBnHL by averaging different number of responses. This figure shows that the quality of the recordings improves as the number of responses increases. The main waves of BAER begin to be identified from the average of 200 responses and are easily recognized from 1000 responses. This fact suggests that 1500 is a suitable number of averages to be able to distinguish the main waves of BAER, reaching a balance between recording time and quality of the recordings.

Figure 13. Brainstem auditory evoked potentials. Influence of the number of biological responses considered on average BAER analysis from a group of subjects Figure 14 displays BAER at different intensities of stimulation recorded over eight normal

hearing adults of both sexes, aged between 23 and 33. Intensities of stimulation involved on this study are 10 dBnHL, 30 dBnHL, 50 dBnHL and 70 dBnHL. Interstimulus interval was set on 25 milliseconds. These signals have been obtained averaging 4000 responses in order to increase the quality of the recordings. Figure 14 highlights the variability of the evoked potentials related to amplitude, latency and morphology of the waves. These differences among subjects are normal [9]. Changes on intensity of stimulation affect amplitude and latency of the peaks, as shown on figure 14. A reduction on the intensity of stimulation produces a decrease in the amplitude of all components of auditory evoked potentials, especially on wave I. The decrease on the amplitude of wave V occurs more slowly, a fact that is used to find the hearing threshold of a subject. The decrease on intensity of stimulation also causes an increase on the latencies of the waves. The latencies of the main waves of the evoked potentials of figure 14 are presented on table 3. Wave V can easily be identified at 70, 50 and 30 dBnHL; but on some subjects, wave III does not appear at low levels. The standard deviation (table 3) shows that the variability of latencies among subjects is small. The mean values on the same table expose that latencies increase as intensity of stimulation decrease. The influence that intensity of stimulation has over latencies of BAER waves has been analyzed through a statistical hypothesis test. Latencies of waves III, V and interpeak latency III-V at 70

Table 3. Mean, standard deviation and p value of latencies III, V and interpeak latency III-V Intensity (dBnHL) Subject 1 Subject 2 Subject 3 Subject 4 Subject 5 Subject 6 Subject 7 Subject 8 Mean Standard deviation p - value 884

70 3.88 4.00 4.04 4.04 3.90 4.08 3.88 3.68 3.94 0.13 -

Lat III (ms) 50 4.16 4.50 4.68 4.06 4.28 4.28 3.96 4.27 0.25 0.0053

30 4.88 4.88 5.32 4.76 5.12 5.12 5.01 0.20 5·10-8

70 5.76 6.04 5.96 6.20 5.88 6.00 5.98 5.64 5.93 0.17 -

Lat V (ms) 50 6.12 6.24 6.20 6.48 6.12 6.20 6.20 5.92 6.18 0.16 0.0083

30 6.88 7.00 6.76 7.30 6.72 7.00 7.06 7.16 6.98 0.19 2·10-8

70 1.88 2.04 1.92 2.16 1.98 1.92 2.10 1.96 1.99 0.09 -

Lat III-V (ms) 50 30 1.96 2.00 1.70 1.88 1.80 1.98 2.06 1.96 1.92 1.88 1.92 1.94 1.96 1.90 1.94 0.12 0.05 0.12 0.23

Volume 6 / Number 4 / 2011


dBnHL have been chosen as reference. The significance level has been set on 0.05, according to many biostatistics applications [10]. The p-values shown on table 3 indicate that intensity of stimulation is a conditioning factor that influences the latency of waves III and V. On the other hand, there is no clear evidence that intensity of stimulation affect the interpeak latency III-V. Several studies support this result [11-12].

Figure 14. Brainstem auditory evoked potentials recorded over eight normal hearing adults. Influence of intensity level of stimulation.


4. Educational approach The use of new technologies in education is a transversal competence that is being promoted by the application of the European Space for Higher Education of the Bologna Declaration. The system described in this paper aims to immerse students into a real-world laboratory exercise with a significant impact on their education. The development of the system for BAER acquisition described on this paper is especially addressed to technical profile students with an approach to biomedical applications. A basic knowledge on the fields of audiology, electronics, signal processing theory and programming skills in MATLAB is required as background education. If necessary, an overview on any of these matters should be provided to students before starting the lab activities. This laboratory suits perfectly on last years of degrees in electronics engineering, computer sciences, biomedical engineering and telecommunication engineering. This laboratory may also fit well in specific masters such as bioengineering and medical physics, multimedia systems, neuroscience, biomedical instrumentation and Ear, Nose and Throat (ENT) studies. This laboratory will help students gain specific skills and competences depending on their field of study. Furthermore, they will have the perspective of being involved in a multidisciplinary project. Electronics engineering students are expected to (a) design and implement a low noise & high gain instrumental amplifier; (b) calibrate the system and (c) characterize the amplifier in terms of its frequency behavior, lin885


earity, consumption, noise spectral density and noise power as a function of the impedance of the source. These activities will help them improve their ability on the field of biomedical instrumentation, learn to specify the performance of a lineal system through its electronic characterization, understand the importance of the input and output impedances of the electronic modules and the relevance of an adequate treatment of noise. Moreover, electronics engineering students will also start to be familiar to analog and digital advanced signal processing techniques and the process of biological signals recording. The activities of telecommunication engineering students will be focused on (a) signal acquisition; (b) configuring of the audio settings on the process of AD/DA conversion performed by a sound card, such as the sampling rate or the number of bits of quantization; (c) the estimation of the quality of a biological recording in terms of SNR; (d) digital filtering of the electroencephalogram; (e) the averaging of biological responses and (f) the use of advanced artifact rejection techniques. These students will also interact with the electronics of the system and will realize that their work contributes to the development of a biomedical application. Computer sciences students will program the digital signal processing routines designed by telecommunication engineering students that are necessary for recording BAER. On these laboratory sessions, students will make progress on their programming skills using MATLAB, learning new functions useful for other disciplines. Furthermore, computer sciences students will design and program a Graphical User Interface, so that medical doctors and other users of non-technical education profile are able to record auditory evoked potentials. The development of these laboratory activities will make computer sciences students be more aware of the relevance of electronics and signal processing on an application with a clinical purpose. The main role of students of audiology is the recording and the analysis of brainstem auditory evoked potentials. Audiology students will apprehend the influence that factors like stimulation intensity or stimulation rate has over the amplitudes and latencies of BAER. In general terms, these students will gain a better understanding in audiology, and more specifically, on the hearing process. These 886

students are expected to perform statistical hypothesis tests to propose biological conclusions and give a physiologic interpretation of the results. This activity will help students enlarge their experience in statistical tests to reach conclusions with a certain confidence interval. This laboratory will also make students understand the process of BAER recording in detail and be more aware of (a) the importance of electronics, (b) the importance of a correct electrodes placement in order to reduce the output impedance of the signal source, (c) the large number of possibilities and new functionalities that the digital signal processing offers and (d) the importance of a friendly interactive user interface. The assessment of the system is also a laboratory activity with a high educational approach. It can be performed not only by recording brainstem auditory evoked potentials from a subject, but also simulating the recording of a digitally synthesized signal with similar shape to a biopotential, i.e. a pseudopotential. The successful recording of a low-amplitude pseudopotential is a challenging laboratory activity. This activity requires no electrodes placement and gives students a high degree of freedom: (a) the shape of the pseudopotential can be digitally synthesized in MATLAB; (b) the amplitude of the pseudopotential can be controlled by the output level of the signal and by the voltage divider; (c) extra noise can be added to the signal by increasing the output impedance Zout or by digital means, where users can also control its spectral distribution; and (d) other parameters like the sampling rate, the number of bits of quantization and the number of responses considered for averaging can also be controlled. The pedagogical methodology proposed to achieve the goals of this lab is experiential learning. This didactic strategy is founded on the gain of knowledge directly from experience [8]. It is well known that experiential learning activities develop a better understanding of a subject, enhance retention and improve cognitive skills [13]. The mission of teachers is to help students organize their ideas, schedule the milestones to complete, lead them to solve specific technical problems and evaluate their results and motivation. This way of learning creates an educational environment that encourages students to be committed with this experience. Volume 6 / Number 4 / 2011


The system described on this paper can be used as an educational tool. Furthermore, because of its open nature, it can also be used for research to investigate the process of hearing and other advanced auditory features. Indeed, this system has already been successfully used obtaining BAER at high rates of stimulation [14]. Assessment of the proposed educational approach A laboratory activity was performed on the Master of Multimedia Technologies framework at the University of Granada (Spain). This master provides education in digital processing methods of audio and video signals, audiology, the development of man-machine interfaces, etc. Students enrolled on this master come usually from technical profile degrees such as computer sciences, electronics engineering and telecommunication engineering. The laboratory session was held after twelve hours of lectures in audiology in which students learnt about anatomy and physiognomy of the auditory system, the process of recording otoacoustic emissions and brainstem auditory evoked potentials, the use of auditory evoked potentials for clinical diagnosis, etc. The aim of the laboratory was to involve the students in a BAER recording session and help them consolidate the most important concepts studied on lectures. The lab comprised the performance of four activities. Activity 1 consisted of a short introduction in which the elements of a BAER recording system were presented and the types of noise that affect the recording of BAER were theoretically analyzed. Activity 2 entailed the placement of the electrodes on the head of a subject and the connection of the hardware elements of the BAER recording system: laptop, external AD/DA sound card, headphones and amplifier. Activity 3 involved the recording of biological responses associated to 2000 acoustic stimuli and the presentation of BAER at different number of averaged responses: 50, 100, 200, 500, 1000, 1500 and 2000. Activity 4 performed the recording of auditory evoked potentials at the stimulation levels: 10 dBnHL, 30 dBnHL, 50 dBnHL and 70 dBnHL. Nine students participated on this laboratory. Students were asked to answer a test of knowledge before and after the laboratory session. This Volume 6 / Number 4 / 2011

practice pursued a double intention: on one hand, the contents learned by the students were evaluated before and after the laboratory; on the other, the initial test prepared the students for the kind of education expected to receive on the laboratory session, helping them focus their attention on the most important contents and therefore, increasing their motivation and interest. The questions that composed the test of knowledge are presented on table 4. Figure 15 shows the percentage of correct answers to the questions of the test of knowledge before and after the laboratory session. Partly correct answered questions count half of the value of a correct answer. Figure 15 shows that students have given more correct answers to the questions of this test after the performance of the laboratory activities than before. Every question has been correctly answered by more than 70 % of the students on the test after the laboratory session. Table 4. Questions of the test of knowledge. Tag Q1 Q2 Q3 Q4 Q5

Question What are the main elements involved in a standard BAER recording system? What is the correct placement of electrodes on the head? What is the methodology used to reduce the effect of non-synchronous artifact? What types of artifact commonly affect the recording of BAER? What is the influence of the intensity level on the amplitude and latency of the waves of auditory evoked potentials?

Figure 15. Test of knowledge results. This figure shows that the percentage of correctly answered questions increase after the laboratory session. 887


Finally, students were asked to complete a satisfaction survey in order to evaluate the efficiency of the laboratory. This evaluation was performed by the measurement of the level of agreement of the students to different statements. Statements and results of this survey are presented on table 5. Level 1 corresponds to “I strongly disagree” and level 5 to “I strongly agree”. Column “evaluation” on table 5 shows the mean and standard deviation of the evaluations of the students. In general terms, table 5 evidences that the laboratory had a high level of acceptance by the students. This table also shows that the students think that the laboratory session was useful to help them understand important concepts studied in lectures, that the laboratory was interesting and that they would recommend it to other audiology and bioengineering students. Students were also asked about their personal opinion of the laboratory, giving answers such as: “practical exercises favor a better comprehension of theory”, “my participation in this laboratory has helped me understand and remember the things explained in lectures” or “this lab makes easier the understanding of theoretical models studied in class”.

5. Conclusion This paper proposes the implementation and assessment of a flexible, low cost and high performance system for acquisition of Brainstem Auditory Evoked Response (BAER). This system is proposed to be used as a didactic tool for the study of auditory evoked potentials in a biomedical, electronics and signal processing laboratory. The hardware and software elements that compose the recording system are thoroughly presented to describe the process of implementation. In order to verify that our system is capable of recording biological responses, a low amplitude pseudopotential was successfully recorded. The improvement of the quality of the BAER signals as the number of responses increases has been evaluated. In addition, BAER from eight normal hearing subjects were recorded at different levels of stimulation. The influence of the level of stimulation over the latencies of the main waves was statistically evaluated, reaching conclusions consistent with other studies. Although there already exist other medical devices able to record BAER, most of them are expensive

Table 5. Satisfaction survey results. The evaluation column represents the mean and standard deviation of the level of agreement of the students to each statement. Level 1 corresponds to “I strongly disagree” and level 5 to “I strongly agree” Statement This laboratory helped me understand the overall performance of a standard BAER recording system Thanks to this lab, I learned the correct placement of the electrodes on the head to record BAER. This laboratory helped me understand the different types of artifact involved in the recording of evoked potentials, as well as the digital signal processing methods used to reduce their effect. I grasped the way in which the number of averaged auditory responses affects the quality of the recorded BAER. This laboratory helped me assimilate the effect that the level of intensity has on the amplitudes and latencies of auditory evoked potentials. In general terms, the contents of this laboratory were presented in a suitable and systematic way. This laboratory session was interesting. The level of difficulty of this lab was appropriate. My motivation and interest in the subject has increased after this lab. I recommend this laboratory to other audiology and bioengineering students I consider that the concepts entrenched on this lab are useful for other audiology and bioengineering students. 888

Evaluation Mean

Std. Dev.

4.6

0.5

4.7

0.7

4.0

0.7

4.3

0.7

4.3

0.7

4.1

0.6

3.9 4.6 3.5 4.4

0.9 0.5 0.5 0.7

4.3

0.5



and suffer from a lack of flexibility, making them inappropriate for education purposes. The laboratory we propose is especially addressed to technical profile students from degrees such as electronics engineering, telecommunication engineering, computer sciences and audiology studies. The development of the activities that compose this BAER recording system in a signal processing laboratory presents several educational advantages. The pedagogical model proposed to be followed is experiential learning, in which students learn basically from their own direct experience. This approach intends to help them comprehend complex concepts and improve retention. A laboratory session was performed on the Master in Multimedia Technologies framework at the University of Granada. This laboratory session comprised the performance of several activities that, according to the self-evaluation given by the students, helped them understand, assimilate and consolidate theoretical concepts studied in lectures. One of the contributions of this article is to provide the necessary knowledge to implement and evaluate a high performance, low-cost and flexible BAER recording system. The results of this work suggest the use of the described BAER recording system as a didactic tool in an audiology, electronics and signal processing laboratory framework because of its lowcost and open nature. The laboratory experience and the analysis of the satisfaction survey results show the educational utility of our system and confirm that an educational strategy based on experiential learning contributes positively to the comprehension and consolidation of knowledge by the students. Acknowledgements This work has been supported by the research project “Diseño, implementacion y evaluacion de un sistema avanzado de registro de potenciales evocados auditivos del tronco (PEAT) basado en señalizacion codificada” (TEC2009-14245), Plan Nacional de I+D+i (2008-2011), Ministry of Science and Innovation (Government of Spain) and European Regional Development Fund Programme (20072013); and by the grant “Programa de Formacion de Profesorado Universitario” (FPU) (AP2009-3150), Ministry of Education, Government of Spain. Authors also thank the collaboration of the subjects and students that have participated in this experience. Volume 6 / Number 4 / 2011

References 1. Hall JW. Handbook of Auditory Evoked Responses, Allyn and Bacon, Boston, 1992. 2. Jewett DL, Williston JS. Auditory evoked far fields averaged from the scalp of humans. Brain. 1971; 4:681-696. 3. Leung S, Slaven A, Thornton ARD, Brickley GJ. The use of high stimulation rate auditory brainstem responses in the estimation of hearing threshold”. Hearing research. 1998; 123(1):201-205. 4. Ruiz JM. Potenciales del tronco cerebral evocados mediante estimulacion electrica en pacientes con implante coclear. PhD Thesis, University of Granada, 2002. 5. Bahmer A, Peter O, Baumann U. Recording of electrically evoked auditory brainstem respones (e-abr) with an integrated stimulus generator in matlab. Journal of Neuroscience Methods. 2008. 173:306-314. 6. Bahmer A, Peter O, Baumann U. Recording and analysis of electrically evoked compound action potentials (ecaps) with med-el cochlear implants and different artifact reduction strategies in matlab. Journal of Neuroscience Methods. 2010. 191:66-74. 7. Aliane N. A MATLAB/Simulink-based interactive module for servo systems learning. IEEE Trans. Educ. 2010. 53(2):165-271. 8. Kolb DA. Experiential learning: Experience as the source of learning and development. Englewood Cliffs, NJ: Prentice-Hall, 1984. 9. Stockard JE, Stockard JJ, Westmoreland BF, Corfits JL. Brainstem auditory evoked responses: normal variation as a function of stimulus and subject characteristics. Arch. Neurol. 1979; 736:823-831. 10. Altman D. Practical statistics for medical research. 1st Ed., London: Chapman & Hall, 1997. 11. Starr A, Achor LJ. Auditory brainstem responses in neurological diseases. Arch. Neurol. 1975; 32:761-768. 12. Rosenhamer HJ, Lindstrom B, Lundborg T. On the use of click-evoked electric brainstem responses in audiological diagnosis. Scand. Audiol. 1978; 7:193. 13. Woods DR, Hrymak AN, Stice JM. The future of engineering education III. Developing critical skills. Chem Eng. Educ. 2000. 34(2):108-117. 14. Alvarez I, Valderrama JT, DeLaTorre A, Segura JC, Sainz M, Vargas JL. Reduccion del tiempo de exploracion de potenciales evocados auditivos del tronco cerebral mediante estimulacion aleatorizada. XXV Simposium Nacional. URSI, 2010.

Corresponding author Joaquin Tomas Valderrama Valenzuela, Dpt. Theory of Signal, Networking and Communications, CITIC-UGR, University of Granada, Spain, E-mail: [email protected] 889

Recording of auditory brainstem response at high stimulation rates using randomized stimulation and averaging Joaquin T. Valderrama,a) Isaac Alvarez, Angel de la Torre, and Jose Carlos Segura Department of Signal Theory, Telematics and Communications, CITIC-UGR, University of Granada, 18071 Granada, Spain

Manuel Sainzb) and Jose Luis Vargas San Cecilio University Hospital, ENT Service, 18071 Granada, Spain

(Received 3 April 2012; revised 3 October 2012; accepted 11 October 2012) The recording of auditory brainstem response (ABR) at high stimulation rates is of great interest in audiology. It allows a more accurate diagnosis of certain pathologies at an early stage and the study of different mechanisms of adaptation. This paper proposes a methodology, which we will refer to as randomized stimulation and averaging (RSA) that allows the recording of ABR at high stimulation rates using jittered stimuli. The proposed method has been compared with quasi-periodic sequence deconvolution (QSD) and conventional (CONV) stimulation methodologies. Experimental results show that RSA provides a quality in ABR recordings similar to that of QSD and CONV. Compared with CONV, RSA presents the advantage of being able to record ABR at rates higher than 100 Hz. Compared with QSD, the formulation of RSA is simpler and allows more flexibility on the design of the pseudorandom sequence. The feasibility of the RSA methodology is validated by an analysis of the morphology, amplitudes, and latencies of the most important waves in ABR recorded at C 2012 Acoustical Society of America. high stimulation rates from eight normal hearing subjects. V [http://dx.doi.org/10.1121/1.4764511]

I. INTRODUCTION

The objective evaluation of hearing is currently a widely used practice in hospitals and clinics around the world. A universal newborn hearing screening is compulsory in most of the United States and it is recommended in Europe (Grandori and Lutman, 1999; American Academy of Pediatrics, 1995). Auditory brainstem responses (ABRs), along with otoacoustic emissions, are objective measurements commonly applied for hearing screening (Erenberg et al., 1999; Kennedy et al., 1991). ABR signals represent the electrical activity of the brainstem associated with an auditory stimulus. This biological response is described by a series of waves that occur during the first 10 ms after the stimulus. These waves are identified with Roman numerals as proposed in Jewett and Williston (1971). The study of ABR is of great interest from an audiological point of view, since it allows the analysis of some of the mechanisms involved in the process of hearing (e.g., Hall, 2007; Katz, 1994). The methodology for an ABR recording consists of the presentation of auditory stimuli to the subject and the recording of their associated electrical response by electrodes placed on the skin in different places of the head. The low amplitude of these potentials (usually less than 1 lV at the electrodes) requires a high amplification in the recording process. Additionally, the signals are contaminated by

a)

Author to whom correspondence should be addressed. Electronic mail: [email protected] b) Also at: Department of Surgery and its Specialties, University of Granada, 18071, Granada, Spain. 3856

J. Acoust. Soc. Am. 132 (6), December 2012

Pages: 3856–3865

artifacts of diverse origin, such as neuromuscular activity of the subject, noise from the amplifier, electromagnetic interference, etc. In order to reduce these artifacts, the response to a large number of stimuli is recorded. If the response to each stimulus can be assumed to be linear and time invariant (LTI), the signal-to-noise ratio (SNR) improves by averaging a large number of sweeps (e.g., Wong and Bickford, 1980; Webster and Clark, 1995). The conventional (CONV) technique for ABR recording consists of the averaging of several sweeps whose corresponding stimuli are periodically presented, i.e., with a constant inter-stimulus interval (ISI). This technique has the important limitation that the ISI must be greater than the averaging window in order to avoid the contamination of the recording by an adjacent response (e.g., Zollner et al., 1976; Kjaer, 1980). Therefore, the CONV technique cannot be used to record ABR at rates higher than 100 Hz. The recording of ABR at high stimulation rates (higher than 80 Hz) may present a number of advantages, as reported by several authors. Thornton and Slaven (1993) and Leung et al. (1998) argue that the presentation of stimuli with a low ISI could reduce the necessary recording time, which is a critical parameter in certain situations such as investigating children and other non-cooperative subjects (e.g., Burkard et al., 1990; Bell et al., 2001). Several authors agree that the use of high stimulation rates would allow a more detailed study of the phenomenon of adaptation (e.g., Lasky, 1997; Burkard et al., 1990). Leung et al. (1998) states that the use of high stimulation rates may help to improve accuracy in estimating the hearing threshold of a subject. Finally, many researchers have found that the use of high stimulation rates

0001-4966/2012/132(6)/3856/10/$30.00

C 2012 Acoustical Society of America V

Author's complimentary copy

PACS number(s): 43.64.Ri, 43.64.Yp [BLM]

J. Acoust. Soc. Am., Vol. 132, No. 6, December 2012

rates (e.g., Bohorquez et al., 2009; Millan et al., 2006). The main advantage of this method with respect to MLS relies on a significant relaxing of the stimulation sequences restrictions. However, CLAD presents two important limitations: The convolution matrix, generated from the stimulation sequence, must be invertible (i.e., not singular) and the power spectral density of the stimulation sequence must accomplish the noise attenuation requirements exposed in the QSD model. Indeed, CLAD is a methodology that can be used to implement QSD (Jewett et al., 2004). The purpose of the present study is to present a new methodology that allows the recording of ABR signals at high stimulation rates using jittered stimulation sequences. We have called this methodology randomized stimulation and averaging (RSA). The RSA technique consists of the averaging of auditory responses corresponding to stimuli whose ISI varies randomly according to a predefined probability distribution. This method [for which fundamentals were described in Alvarez et al. (2010) including some preliminary results] is described in detail in this paper and compared with QSD and CONV in terms of quality of the responses. ABR responses obtained with RSA present a similar quality as those with QSD and CONV. In comparison with CONV, RSA allows the recording of ABR at rates higher than 100 Hz; furthermore, the restrictions imposed to QSD, CLAD, or MLS sequences are not necessary in RSA, which makes the implementation of RSA easier and allows more flexibility in the selection of the probability distribution of the ISI. This allows more control in the jitter of the sequences used for RSA. The reliability of the RSA technique has also been assessed by an analysis of the morphology, amplitudes, and latencies of the main waves in ABR recorded from eight subjects at different stimulation rates, obtaining results consistent with previous studies. II. RSA

In contrast to the CONV stimulation technique, in which stimuli are presented at a constant period greater than the averaging window [Fig. 1(A)], the RSA technique consists of averaging auditory responses, corresponding to a burst of stimulation pulses, in which the ISI varies randomly according to a predefined probability distribution. The RSA technique involves a digital blanking process and non-uniform averaging in order to minimize the effect of the stimulus artifact in overlapped responses. The digital blanking process consists of considering null values those samples of the EEG in which stimulus artifact occurs. Figure 1(B) shows a frame of a RSA stimulation signal of ISI4–8, in which the ISI randomly varies according to a uniform distribution between 4 and 8 ms. Figure 1(C) shows a histogram of the ISI for the selected random sequence. When RSA is applied, two important differences can be observed with respect to CONV stimulation: The rate of stimulation can be higher and two or more consecutive responses can be overlapped; both effects depend on the selected probability distribution of the ISI. In order to obtain the ABR response in the RSA framework, a non-uniform averaging is applied to the raw electroencephalogram (EEG), after a digital blanking process. Let Valderrama et al.: Randomized stimulation and averaging

3857


could be useful for detecting certain pathologies at an early stage (e.g., Don et al., 1977; Stockard et al., 1978; Yagi and Kaga, 1979; Jiang et al., 2000; Thornton et al., 2006). Several techniques which are able to record ABR at high rates of stimulation have emerged with the intention of breaking with the limitation imposed by the CONV technique. These methods estimate the biological response through the use of pseudorandom stimulation sequences that are repeated periodically. The jitter of a stimulation sequence is a measure that indicates the magnitude of dispersion of the ISI. All techniques only recover the LTI-like component of the total response, while ignoring deviations from LTI behavior. However, as the stimulation rate increases, the morphology of responses changes (Bohorquez and Ozdamar, 2006). Therefore, the jitter of a stimulation sequence could be a critical parameter to be considered when assuming that each click evokes the same response (e.g., Jewett et al., 2004). The most relevant techniques proposed for recording ABR at high stimulation rates are maximum length sequences (MLS), quasiperiodic sequence deconvolution (QSD), and continuous loop averaging deconvolution (CLAD). The MLS technique was developed by Eysholdt and Schreiner (1982). In this technique, the deconvolution of overlapping responses is performed using bursts of pseudorandom pulses whose ISIs are adjusted to De-Brujin sequences. This technique has not only been used to record ABRs at high stimulation rates (e.g., Leung et al., 1998; Bohorquez and Ozdamar, 2006) but also to record mid-latency auditory evoked potentials (e.g., Lavoie et al., 2010) and other biological signals such as transient evoked otoacoustic emissions (e.g., Hine et al., 2001; Thornton, 1993; de Boer et al., 2007). The stimulation technique based in Legendre sequences (LGS) is also adjusted to the De-Brujin sequences. This methodology has also been used to obtain ABRs at high rates of stimulation. MLS and LGS present a similar performance, as reported by Burkard et al. (1990). The main difficulties with both methodologies are their high jitter and the restrictions imposed on the stimulation sequences. The QSD model, developed by Jewett et al. (2004), describes the conditions that sequences of low jittered pulses have to fulfill in the frequency domain in order to allow deconvolution of overlapping responses at high rates of stimulation. The main difficulty with this technique resides in the search for an optimal stimulation sequence, which usually requires a high computational effort (e.g., Jewett et al., 2004; Wang et al., 2006). The frequency domain analysis described in the QSD model opened a new framework that has influenced the approach of other methodologies. In the CLAD methodology, devised by Ozdamar et al. (2003a,b) and Delgado and Ozdamar (2004), deconvolution of overlapping responses is achieved through time-domain matrix algebra processing. A frequency domain formulation of this technique was described by Ozdamar and Bohorquez (2006). The Wiener filtering theory can be applied to the CLAD methodology to improve the SNR of the recordings (Wang et al., 2006). Although the most important application of the CLAD deconvolution method is the recording of ABR, it has also been used to record electrocochleograms and auditory middle latency responses at high stimulation

ABR response when RSA is applied. A real ABR response (obtained with 10 000 sweeps) was used to artificially synthesize an EEG. A stimulation sequence was generated at ISI8–12 (i.e., with a uniform distribution between 8 and 12 ms). Figures 2(A) and 2(B) show, respectively, the beginning of the stimulation signal and the synchronization signal s(n). The synchronization signal s(n) was convolved with the ABR response and white noise was added at SNR ¼ 13 dB in order to obtain the synthesized EEG y(n) [Fig. 2(C)]. Figure 2(D) shows the effect of the digital blanking process, y(n) b(n). Those segments of the EEG with a stimulus artifact are removed from averaging. Finally, Fig. 2(E) shows the ABR response x^ðjÞ averaged according to the RSA technique. The sampling rate in this example was fs ¼ 25 kHz. Although in ABR recordings the EEGs present a typical SNR below 10 dB, in this example the synthesized EEG was contaminated with less noise in order to allow the identification of biological responses. In contrast to the rest of the techniques able to record ABR at high stimulation rates, RSA does not perform deconvolution. In the RSA methodology, there basically exist three types of artifacts involved in the process of ABR FIG. 1. (A) Example of a stimulation signal in CONV mode. Stimuli are presented at a constant period of 22 ms, greater than the averaging window (Aw ¼ 10 ms). (B) Example of stimulation signal in the RSA technique at ISI4–8. In this case, ISI is smaller than the averaging window. (C) Histogram of the ISI when a uniform distribution between 4 and 8 ms is considered to generate the sequence (ISI4–8).

y(n), b(n), and s(n) (n ¼ 1,…,N) be, respectively, the digitized EEG, the blanking signal, and the synchronization signal (that indicates with the value of 1 the samples in which stimulation starts and 0 otherwise). For an EEG in which K stimuli are presented, the index of the samples in which each stimulus starts can be represented with m(k) (k ¼ 1,…,K). Therefore, s(m(k)) ¼ 1. The blanking signal b(n) differences valid samples of EEG (value 1) from samples contaminated with a stimulation artifact (value 0). The digital blanking process considers null values 0.2 ms before and 0.8 ms after each stimulus [Eq. (1)]. The ABR signal x^ðjÞ is estimated in RSA by averaging of the biological responses corresponding to the K stimuli presented to the subject, without considering in the averaging process those segments of EEG affected by stimulation artifact [Eq. (2)] 8 < 0 if n 2 ½mðkÞ 0:2 ms fs ; n ¼ 1; …; N; mðkÞ þ 0:85 ms fs ; 8k bðnÞ ¼ : 1 otherwise; (1)

x^ðjÞ ¼

k¼1 K X bðmðkÞ þ jÞ

;

j ¼ 1; …; J;

k¼1

(2) where J and N represent, respectively, the length of the averaging window and the total number of samples of the EEG. Figure 2 shows an example to illustrate the calculation of the 3858


FIG. 2. An illustration of the estimation of the ABR response based on RSA. (A) RSA stimulation signal for ISI8–12. (B) Synchronization signal s(n). (C) Raw EEG y(n) (in this example, the EEG was synthesized from a real ABR response). (D) Effect of digital blanking, b(n) y(n). Segments of EEG with stimulation artifacts removed from averaging. (E) Estimation of the ABR response x^K ðjÞ. Sampling frequency: fs ¼ 25 kHz; length of the averaging window: J ¼ 250 samples (10 ms). Valderrama et al.: Randomized stimulation and averaging


K X bðmðkÞ þ jÞ yðmðkÞ þ jÞ

III. METHODS A. Acquisition of EEG

The evaluation of the RSA method was based on the analysis of ABR recorded from different subjects. The stimulation of the auditory system was performed by 0.1 ms clicks presented at an intensity of 70 dBnHL. Zero dBnHL was established considering the threshold level (stimulation level at which the stimulus is just detectable) for a burst of clicks presented at a rate of 20 per second in a group of 15 subjects (9 male and 6 female) ranging in age between 24 and 31 yr, with no self-reported history of auditory dysfunction (normal hearing subjects). The equivalent 0 dBnHL under such typical stimulus conditions is 36.4 dB peak sound pressure level and 29.9 dB peak-to-peak equivalent sound pressure level (Burkard, 1984; Klein and Teas, 1978; Stapells et al., 1982). Duration clicks of 0.1 ms were used as stimuli in order to evoke synchronous firing of a large number of neurons, especially those in the 1000 to 4000 Hz region (Hall, 2007). The recording sessions were held in a room isolated from acoustical and electromagnetic interferences. During the process of ABR recording, subjects were seated in a comfortable and relaxed position to minimize electromyogenic noise. Auditory stimuli were presented to the subjects through standard circumaural headphones (Pro-550, Ultrasone, Wielenbach, Germany). The biological evoked responses associated with the stimuli were recorded from three Ag/AgCl surface electrodes placed on the skin at different positions on the head. Active, reference, and ground electrodes were situated at the high forehead, the ipsilateral mastoid, and on the low forehead, respectively. Interelectrode impedances were always below 10 kX. The biological signal was 70 dB amplified and band pass filtered (100 to 3500 Hz). Both the biological signal and the stimulation signal were sampled at 25 kHz and represented with 16 bits/sample (with a two-channel analog-to-digital conJ. Acoust. Soc. Am., Vol. 132, No. 6, December 2012

verter). Digital signals were processed with algorithms implemented in MATLAB (The Mathworks, Inc., Natick, MA). The recorded EEG had been digital filtered using a sixth order bandpass Butterworth filter (150 to 3000 Hz). The synchronization signal s(n) had been obtained from the recorded stimulation signal. A full description of the ABR recording system can be found in Valderrama et al. (2011). B. Recording of RSA, QSD, and CONV responses

This study involves the recording of ABR signals over a group of 8 normal hearing adults (5 males and 3 females; aged between 22 and 36 yr) using the RSA, QSD, and CONV techniques. These subjects were volunteers and were informed in detail about the experimental protocol and possible side effects of the test. A jitter of 4 ms was used in the stimulation sequences of RSA and QSD methodologies. Sweeps (20 000) were recorded from each subject using the RSA and QSD techniques at the stimulation rates corresponding to ISI20–24, ISI16–20, ISI12–16, ISI10–14, ISI8–12, ISI6–10, ISI4–8, and ISI2–6. The same number of sweeps were recorded using the CONV technique at the stimulation rates ISI22, ISI18, ISI14, ISI12, and ISI10. The RSA stimulation sequences were randomly generated with a uniform probability distribution of the ISI between two limits. RSA responses were obtained according to the procedure exposed in Sec. II. In this work the stimulation sequences in QSD (q-sequences) are composed of 16 stimuli with ISI in a range between two limits. These sequences are periodically repeated to provide the required number of sweeps. In order to obtain efficient q-sequences, for each ISI condition 3 million sequences with ISI in the specified range were randomly generated, and the magnitude of the fast Fourier transform (FFT) was evaluated for each sequence. The selected q-sequence was the one which maximizes the minimum value of the FFT magnitude in the band of interest (150 to 3000 Hz). This way the selected q-sequence minimizes the amplification of noise when deconvolution is applied to obtain the ABR responses. QSD responses corresponding to each block of 16 stimuli were averaged. The final QSD response was obtained by deconvolution and filtering of the averaged block of responses. A more detailed description of the QSD methodology can be found in Jewett et al. (2004). A basic artifact rejection procedure has been applied to RSA, QSD, and CONV recordings. In RSA and CONV, the response to each stimulus was evaluated after the digital blanking process and those sweeps whose instantaneous voltage exceeded 10 lV were removed from averaging. In the case of QSD, the analysis was similar but estimation of the instantaneous voltage is referred to each block of 16 stimuli. The blocks with instantaneous voltage higher than 10 lV after digital blanking were not considered for deconvolution or averaging. This way, the artifact rejection procedure removes the noisiest parts of the EEG from the final ABR response. C. Quality assessment of ABR responses

This article evaluates and compares the performance of the RSA, QSD, and CONV techniques. An objective assessment of the quality of recordings is performed in this study Valderrama et al.: Randomized stimulation and averaging

3859


recording: Stimulus artifact, noise unlocked with stimuli, and noise associated with overlapping responses. The digital blanking and averaging processes minimize, respectively, the effect of the first two types of artifacts. The noise associated with overlapping responses can be minimized by averaging and an adequate selection of the jitter in the stimulation sequence. The RSA technique is able to retrieve the ABR to a single click as long as the jitter of the stimulation sequence is large enough. However, very high jittered stimulation sequences could lead to obtain inaccurate ABR signals since auditory responses of different morphology would be averaged, and therefore, considered the same evoked response. The jitter of a stimulation sequence should be chosen carefully, reaching a compromise between these two aspects. Moreover, since the use of the digital blanking technique produces a non-uniform number of averaged responses inside the averaging window, we consider that an adequate jitter should allow a number of averaged responses over a threshold of the 70% of the total number of available responses all along the averaging window. Averaging a number of responses below such threshold could produce a noticeable difference in terms of quality between different segments of the response.

D. ABR amplitudes and latencies

This paper presents an analysis of the amplitudes and latencies of the main waves of ABR signals obtained using the RSA, QSD, and CONV techniques at different stimulation rates over a group of eight normal hearing subjects. The latencies and amplitudes of waves I, III, and V were measured on ABR recordings obtained using 20 000 sweeps, applying the RSA and QSD techniques at the stimulation rates ISI20–24, ISI16–20, ISI12–16, ISI10–14, ISI8–12, ISI6–10, ISI4–8, and ISI2–6, and using the CONV technique at the stimulation rates ISI22, ISI18, ISI14, ISI12, and ISI10. Latencies were measured as a difference in milliseconds between the stimulus onset and the top of the peaks. Amplitudes were measured in microvolts as a difference between the top of the peak and the following trough for all waves. The mean and standard deviation of the amplitudes and latencies among the eight subjects were calculated at each stimulation rate, and the effect of the stimulation rate over amplitudes and latencies was analyzed by linear regression in the RSA technique. The results obtained in this study were contrasted with previous literature in order to test the efficiency of the RSA technique when recording real ABR signals.

selected sequence in QSD, ABR responses have been compared for different RSA and QSD sequences. Figure 3 shows ABRs from subject 2 involving 10 000 sweeps in the RSA and QSD techniques, using ISI10–14 stimulation signals. Signals A1, A2, and A3 are examples of auditory evoked potentials deconvolved by the QSD methodology using three different efficient q-sequences. These recordings are of high quality. When the former premise is not accomplished and the value of any frequency component is near zero in the passband, the noise at that frequency is increased in the process of deconvolution. Signals B1 and B2 in Fig. 3 are examples of auditory evoked potentials deconvolved by the QSD methodology using two different non-efficient q-sequences. The q-sequences used to deconvolve these signals have a component near 0 at 2.29 kHz (B1) and at 1.64 kHz (B2) and, as a result, the noise at that frequency is significantly increased. Such noise can be identified in the B1 and B2 evoked potentials despite the use of a large number of averaged sweeps. The most important waves can still be distinguished in B1 and B2 signals. However, if the q-sequence had a close to 0 component near 500 Hz, the evoked potentials would not be successfully deconvolved since the main waves of ABR have their major energy around that frequency (Delgado and Ozdamar, 1994). Graphic B3 presents the deconvolved signal using a q-sequence with a very close to 0 component at a frequency of 481.6 Hz. No evoked potentials can be identified in that signal. Signals B1, B2, and B3 show the consequences of non-efficient stimulation sequences in the QSD technique. In the RSA technique, no attention is required to the frequency components of the stimulation sequence since evoked potentials are obtained through an averaging process. Signals C1, C2, and C3 show examples of ABR responses obtained using the RSA technique with different random sequences (from the same subject, at the same stimulation level, and with the same number of sweeps on average). The most important waves

IV. RESULTS A. Selection of the stimulation sequence

The q-sequence in the QSD methodology must be chosen carefully in order to successfully deconvolve the evoked potentials. The power spectral density of an efficient q-sequence avoids frequency components near zero in the passband (Jewett et al., 2004). On the other hand, RSA sequences are not affected by such restrictions. In order to evaluate the importance of the 3860


FIG. 3. Examples of ABR responses obtained using the QSD and RSA methodologies on one subject (subject 2) in the same recording conditions using ISI10–14 stimulation signals. Efficient (A1–A3) and non-efficient (B1– B3) q-sequences are considered in QSD. Responses obtained with 10 000 sweeps. The amplitude of the B3 signal is divided by a factor of 10 to fit onto the graph. While selection of the q-sequence is critical in QSD, no significant differences are observed for different RSA sequences (C1–C3). Valderrama et al.: Randomized stimulation and averaging


through the correlation coefficient (r) methodology. This parameter points out the grade of similarity between two ABR signals. A high positive correlation coefficient would indicate a high quality ABR recording when the two signals are recorded in the same conditions (Mason et al., 1977; Schimmel et al., 1974; Weber and Fletcher, 1980). In comparison with the objective quality assessment methodologies for ABR recordings based on the variance ratio and multiple pre-post Z, the correlation coefficient is considered a more consistent technique to score the quality of ABR recordings (Arnold, 1985). In order to evaluate the performance of the RSA, QSD, and CONV techniques at different stimulation rates, we have split the 20 000 sweeps of each recording into five groups of 4000 sweeps. Then, we have obtained the ABR signals in each group applying the corresponding methodology (RSA, QSD, or CONV). Therefore, 5 different ABR signals of 4000 sweeps, recorded on the same conditions, are obtained from each recording. The correlation coefficient was calculated between all possible combinations of these five ABR signals. Thus, the total number of statistics per recording is ten, that is, combinations of five elements taken two at a time. Finally, the quality of each technique in each stimulation rate can be parameterized with the mean and standard deviation of the statistics of the eight subjects in each scenario.

can be identified in all three recordings and the quality of the recordings is similar for C1, C2, and C3, independent of the random sequence used. B. RSA/QSD/CONV comparison

The performance of the RSA, QSD, and CONV techniques is compared in this section. Each ABR used in this study considered 4000 sweeps in the averaging process. Figure 4 shows examples of ABRs from subject 1 obtained using the RSA, QSD, and CONV techniques at different stimulation rates. Waves I, III, and V can be easily identified in these recordings at all stimulation rates, which suggests that use of such a number of sweeps is appropriate to obtain ABR recordings of enough quality at every stimulation rate considered in this study. Table I presents the results of the quality evaluation test performed over the RSA, QSD, and CONV techniques as a function of stimulation rate. Table I shows the mean and standard deviation of the correlation coefficient (r) calculated between all possible combinations of 5 recordings of 4000 sweeps, taking 2 at a time (i.e., 10 parameters per subject) at each technique and stimulation rate. Since 8 subjects are considered in this study, the total number of parameters in each scenario is 80. Table I demonstrates the great efficiency of the three methods to successfully obtain high quality ABR signals. Table I highlights that the performance of the RSA technique is very similar to CONV but with the advantage of being able to record ABR at rates higher than 100 Hz. Table I also shows that RSA presents a slightly better performance than QSD, especially at high stimulation rates. In general terms, the quality of recordings decreases as the stimulation rate increases in all techniques. This deterioration of clarity of recordings with increasing stimulation rate is a common phenomenon as a consequence of adaptation, as has been reported in previous studies (e.g., Don et al., 1977; Kjaer, 1980; Lasky, 1984). The low standard

TABLE I. Analysis of the correlation coefficient (r) calculated between all possible combinations of 5 recordings of 4000 sweeps, taking two at a time (i.e., 10 parameters per subject) at each technique and stimulation rate. Eight subjects are considered in this study; thus, the total number of parameters in each scenario is 80. Mean (and standard deviation in parentheses) are indicated for each condition. The mean of the rate for each experiment is indicated in the first column. Experiment ISI20–24/22 (45.5 Hz) ISI16–20/18 (55.5 Hz) ISI12–16/14 (71.4 Hz) ISI10–14/12 (83.3 Hz) ISI8–12/10 (100 Hz) ISI6–10/8 (125 Hz) ISI4–8/6 (166.7 Hz) ISI2–6/4 (250 Hz)

RSA

QSD

CONV

0.95 (0.04) 0.93 (0.07) 0.93 (0.06) 0.93 (0.04) 0.89 (0.09) 0.89 (0.09) 0.84 (0.10) 0.80 (0.14)

0.88 (0.08) 0.86 (0.09) 0.80 (0.18) 0.79 (0.18) 0.68 (0.29) 0.69 (0.22) 0.68 (0.20) 0.63 (0.25)

0.95 (0.05) 0.95 (0.05) 0.93 (0.06) 0.90 (0.08) 0.90 (0.08)

deviation in all techniques at low stimulation rates points out a steady measure of quality among recordings in such conditions. The variability of quality increases with the stimulation rate in all techniques. This variation is more remarkable in the QSD technique, where its standard deviation increases to a greater extent than in RSA and CONV, which suggests that QSD is more sensitive to noise than the other two techniques. C. Analysis of amplitudes and latencies measured with RSA

ABR recordings of 20 000 auditory responses were recorded from 8 subjects for this experiment at different stimulation rates using the RSA, QSD, and CONV techniques. Figure 5 shows the recordings corresponding to the RSA methodology. Waves I, III, and V are labeled in Fig. 5. These waves can be easily recognized at most of the stimulation rates, being more difficult the identification of waves I and III


Valderrama et al.: Randomized stimulation and averaging

3861


FIG. 4. Examples of ABR from subject 1 obtained using the RSA, QSD, and CONV techniques at different stimulation rates, considering 4000 auditory responses in the averaging process.

at higher stimulation rates. Wave V was identified in all subjects at all stimulation rates. Waves I and III could be identified at least in six subjects at each stimulation rate, except wave III at the stimulation rate ISI2–6, which could only be 3862


identified in four subjects. The amplitudes and latencies of waves I, III, and V were measured to perform this test. The morphology of the waves, the amplitudes, and the latencies of the most important waves in these recordings are very similar Valderrama et al.: Randomized stimulation and averaging


FIG. 5. ABRs recorded from eight normal hearing subjects. These responses were obtained using the RSA technique at the stimulation rates ISI20–24, ISI16–20, ISI12–16, ISI10–14, ISI8–12, ISI6–10, ISI4–8, and ISI2–6, considering 20 000 sweeps in the averaging process.

with other studies (e.g., Yagi and Kaga, 1979; Lasky, 1984; Lina-Granade et al., 1993; Leung et al., 1998; Jiang et al., 2009; Stone et al., 2009). Table II shows the mean and standard deviation of the latencies and amplitudes of waves I, III, and V on ABR obtained using the RSA, QSD, and CONV techniques at different stimulation rates. A comparison of the mean and standard deviation of these parameters between different techniques indicates that the latencies and amplitudes of the main waves of ABR using the RSA, QSD, and CONV techniques at all stimulation rates are statistically comparable. In addition, the results of the analysis of the amplitudes shown in Table II point out that the amplitude of all waves, on average, decreases as stimulation rate increases. Wave V presents the largest amplitude at all stimulation rates. The large standard deviation of these results points out a significant variability among subjects in terms of amplitudes. The linear regression analysis performed in the RSA technique (AI: r ¼ 0.44, p < 103; AIII: r ¼ 0.49, p < 104; AV: r ¼ 0.56, p < 105) indicates that the stimulation rate is a statistically significant factor that influences the amplitude of ABR signals, as stated in previous literature (e.g., Lasky, 1984, 1997; Leung et al., 1998; Thornton and Slaven, 1993). The analysis of latencies of waves I, III, and V on ABRs obtained at different stimulation rates using the RSA technique is presented in Fig. 6. Figure 6 shows the mean and standard deviation of the latencies of the eight subjects at each stimulation rate. Figure 6 also shows the correlation coefficient (r) and the p-value (probability of the null hypothesis of statistical independence between ISI and latency) of a linear regression analysis of the data. According to this analysis, the latency of wave I is hardly affected by an increase in stimulation rate (r ¼ 0.33; p ¼ 0.018), the

FIG. 6. Latencies of waves I, III, and V recorded using the RSA technique at different stimulation rates. The plots represent the mean values for eight subjects and the error bars represent standard deviations. The correlation coefficient r and the p-value are shown in the plots.

latency of wave III undergoes a slight shift (r ¼ 0.55; p < 105), and the latency of wave V is a deeper shift (r ¼ 0.75; p < 1012). A statistically significant effect is observed in waves III and V. The correlation coefficients are relatively high in the case of waves III and V, and the dependence of latency on ISI is clear (despite the low number of subjects included in the study and the inter-subject variability). This analysis highlights that the stimulation rate influences the central components of the auditory system in a greater extent than peripheral components, as has already been reported in previous studies (e.g., Pratt and Sohmer, 1976; Yagi and Kaga, 1979; Jiang et al., 2009).

TABLE II. Mean (and standard deviation in parentheses) of the latencies (L) measured in ms and amplitudes (A) measured in lV on ABR recorded from eight normal hearing subjects using the RSA, QSD, and CONV methodologies at different stimulation rates, considering 20 000 sweeps in the averaging process. The ABR corresponding to the RSA technique are shown in Fig. 5. QSD

CONV

LI

LIII

LV

LI

LIII

LV

LI

LIII

LV

ISI2024=22 ISI1620=18 ISI1216=14 ISI1014=12 ISI812=10 ISI610 ISI48 ISI26

1.54 (0.20) 1.52 (0.13) 1.56 (0.12) 1.57 (0.13) 1.58 (0.13) 1.69 (0.22) 1.63 (0.16) 1.67 (0.19)

3.74 (0.14) 3.80 (0.13) 3.86 (0.14) 3.90 (0.16) 3.94 (0.17) 4.00 (0.18) 4.03 (0.20) 3.99 (0.09)

5.69 (0.23) 5.79 (0.23) 5.90 (0.23) 5.97 (0.25) 6.07 (0.25) 6.21 (0.22) 6.40 (0.27) 6.72 (0.29)

1.52 (0.15) 1.53 (0.15) 1.57 (0.16) 1.63 (0.16) 1.58 (0.22) 1.64 (0.18) 1.64 (0.19) 1.65 (0.23)

3.79 (0.15) 3.82 (0.13) 3.89 (0.11) 3.94 (0.16) 3.97 (0.15) 4.05 (0.23) 3.98 (0.03) 4.38 (0.23)

5.75 (0.24) 5.83 (0.24) 5.97 (0.25) 6.02 (0.22) 6.15 (0.25) 6.27 (0.25) 6.41 (0.21) 6.73 (0.35)

1.54 (0.16) 1.54 (0.15) 1.57 (0.14) 1.58 (0.14) 1.59 (0.17)

3.77 (0.11) 3.83 (0.16) 3.90 (0.16) 3.90 (0.12) 3.90 (0.09)

5.75 (0.26) 5.83 (0.27) 5.94 (0.26) 6.02 (0.27) 6.04 (0.25)

ISI2024=22 ISI1620=18 ISI1216=14 ISI1014=12 ISI812=10 ISI610 ISI48 ISI26

AI 0.24 (0.09) 0.23 (0.08) 0.23 (0.05) 0.21 (0.05) 0.21 (0.03) 0.17 (0.05) 0.16 (0.03) 0.16 (0.03)

AIII 0.24 (0.08) 0.21 (0.08) 0.19 (0.07) 0.17 (0.09) 0.15 (0.08) 0.15 (0.07) 0.13 (0.06) 0.11 (0.07)

AV 0.28 (0.07) 0.28 (0.09) 0.25 (0.07) 0.24 (0.07) 0.21 (0.07) 0.18 (0.06) 0.18 (0.08) 0.15 (0.03)

AI 0.25 (0.08) 0.23 (0.07) 0.23 (0.06) 0.22 (0.03) 0.19 (0.05) 0.17 (0.05) 0.17 (0.07) 0.15 (0.05)

AIII 0.23 (0.09) 0.20 (0.08) 0.20 (0.10) 0.16 (0.08) 0.15 (0.08) 0.12 (0.04) 0.11 (0.03) 0.10 (0.04)

AV 0.31 (0.08) 0.26 (0.09) 0.23 (0.07) 0.23 (0.05) 0.22 (0.07) 0.21 (0.05) 0.19 (0.05) 0.12 (0.05)

AI 0.24 (0.09) 0.22 (0.07) 0.24 (0.06) 0.16 (0.04) 0.19 (0.09)

AIII 0.23 (0.06) 0.21 (0.08) 0.18 (0.09) 0.19 (0.09) 0.15 (0.06)

AV 0.29 (0.09) 0.28 (0.10) 0.25 (0.08) 0.22 (0.08) 0.21 (0.08)



3863


RSA

This paper describes the RSA technique, a new methodology that can be used to obtain ABR responses evoked by jittered stimuli at high stimulation rates. In this work, the performance of the RSA technique is compared with the QSD and CONV techniques. The search for an optimal stimulation sequence in QSD may accomplish frequency-domain restrictions in order to successfully deconvolve ABR. Otherwise the evoked response would be contaminated by noise in the deconvolution process (Jewett et al., 2004). In contrast to these restrictions, RSA does not impose any frequencydomain constraints on the stimulation sequence, leading to an easier and more flexible implementation of RSA. The quality of ABR responses acquired with RSA has been compared with that corresponding to QSD and CONV, in terms of the correlation coefficient between pairs of ABR signals recorded in similar conditions. This test has been performed at stimulation rates up to 250 Hz (ISI2–6) in the QSD and RSA methodologies, and up to 100 Hz (ISI10) in the CONV technique. The results of this study suggest: (1) That the quality degrades when the ISI decreases (when the stimulation rate increases) because of the reduction of the amplitude of the response; (2) that the quality of ABR signals recorded using RSA and CONV is very similar but with the advantage for RSA of being able to record ABR at rates higher than 100 Hz; and (3) that the quality of the responses recorded with RSA is slightly better than that of the QSD responses, especially at higher stimulation rates. Two mechanisms could be involved in the improvement of RSA with respect to QSD. On one hand, the quality of QSD is strongly influenced by the selected sequence, since noise could be amplified at specific frequencies. In this sense, RSA responses seem to be more stable and independent of the selected stimulation sequence. On the other hand, the procedure selected for artifact rejection has different effects on RSA and QSD: Since QSD responses are obtained from deconvolution of blocks of 16 responses the artifact rejection procedure accepts or rejects each whole block depending on the evaluation of the noise affecting it. However, in the case of RSA, the response to each stimulus can be independently accepted or rejected by the artifact rejection procedure. This results in a more flexible application of the artifact rejection procedure in the case of RSA, since the portions rejected for averaging are smaller in RSA than in QSD. As a consequence, for similar SNRs of the EEG in RSA and QSD recordings, the average SNR of the accepted part of the EEG would be slightly higher in RSA than in QSD, leading to better quality in the resulting ABR responses. Furthermore, RSA provides more accurate ABR signals than CONV at stimulation rates near 100 Hz because the use of a fixed ISI will systematically contaminate the ABR with later components of adjacent responses that are time-locked with the stimulus, and therefore its effect cannot be diminished by averaging (Kjaer, 1980). The jitter of stimulation sequences in RSA or QSD may avoid this undesired effect. A comparison of amplitudes and latencies measured on high quality ABRs obtained using the RSA, QSD, and CONV techniques at different stimulation rates indicates 3864


that ABR recordings obtained with different stimulation techniques are statistically comparable. The RSA technique was also applied to perform an analysis of the influence of the stimulation rate on the amplitudes and latencies of ABRs obtained at different stimulation rates. The results of this analysis are consistent with those reported in previous literature when other methods are applied for recording ABR at high stimulation rates. ACKNOWLEDGMENTS

The authors of this paper acknowledge Professor Dr. A. R. D. Thornton (University Hospital Southampton, NHS Foundation Trust, Royal South Hants Hospital, Southampton, UK) for his review and constructive suggestions on earlier drafts of this manuscript. The authors also thank the collaboration of the subjects that have participated in this study. This work has been supported by the research project “Design, implementation and evaluation of an advanced system for recording ABR based on encoded signaling” (TEC2009-14245), R&D National Plan (2008–2011), Ministry of Economy and Competivity (Government of Spain), and “European Regional Development Fund Programme” (2007–2013), and by the grant “Programa de Formaci on de Profesorado Universitario” (FPU) (AP2009-3150), Ministry of Education, Culture, and Sports (Government of Spain). Alvarez, I., Valderrama, J. T., DeLaTorre, A., Segura, J. C., Sainz, M., and Vargas, J. L. (2010). “Reduccion del tiempo de exploracion de potenciales evocados auditivos del tronco cerebral mediante estimulacion aleatorizada (Brainstem auditory evoked potentials time reduction through randomized stimulation),” XXV Simposium Nacional URSI, p. 4. American Academy of Pediatrics Joint Committee on Infant Hearing 1994 Position Statement (1995). Pediatrics 95, 152–156. Arnold, S. A. (1985). “Objective versus visual detection of the auditory brain stem response,” Ear Hear. 6, 144–150. Bell, S. L., Allen, R., and Lutman, M. E. (2001). “The feasibility of maximum length sequences to reduce acquisition time of the middle latency response,” J. Acoust. Soc. Am. 109, 1073–1081. Bohorquez, J., and Ozdamar, O. (2006). “Signal to noise ratio analysis of maximum length sequence deconvolution of overlapping evoked potentials,” J. Acoust. Soc. Am. 119, 2881–2888. Bohorquez, J., Ozdamar, O., McNeer, R., and Morawski, K. (2009). “Clinical applications of evoked potential Continuous Loop Averaging Deconvolution (CLAD),” in IFMBE Proceedings of the 25th Southern Biomedical Engineering Conference, Vol. 24, pp. 133–134. Burkard, R. (1984). “Sound pressure level measurement and spectral analysis of brief acoustic transients,” Electroencephalogr. Clin. Neurophysiol. 57, 83–91. Burkard, R., Shi, Y., and Hecox, K. E. (1990). “A comparison of maximum length and Legendre sequences for the derivation of brain-stem auditoryevoked responses at rapid rates of stimulation,” J. Acoust. Soc. Am. 87, 1656–1664. de Boer, J., Brennan, S., Lineton, B., Stevens, J., and Thornton, A. R. D. (2007). “Click-evoked otoacoustic emissions (CEOAEs) recorded from neonates under 13 hours old using conventional and maximum length sequence (MLS) stimulation,” Hear. Res. 233, 86–96. Delgado, R. E., and Ozdamar, O. (1994). “Automated auditory brainstem response interpretation,” IEEE Eng. Med. Biol. Mag. 13, 227–237. Delgado, R. E., and Ozdamar, O. (2004). “Deconvolution of evoked responses obtained at high stimulus rates,” J. Acoust. Soc. Am. 115, 1242–1251. Don, M., Allen, A., and Starr, A. (1977). “Effect of click rate on the latency of auditory brain stem responses in humans,” Ann. Otol. Rhinol. Laryngol. 86, 186–195. Erenberg, A., Lemons, J., Sia, C., Tunkel, D., Ziring, P., Adams, M., Hostrum, J., McPherson, M., Paneth, N., and Strickland, B. (1999). “Newborn Valderrama et al.: Randomized stimulation and averaging


V. SUMMARY AND CONCLUSIONS


28th Annual International Conference of the IEEE, Engineering in Medicine and Biology Society EMBS’06, pp. 4987–4990. Ozdamar, O., and Bohorquez, J. (2006). “Signal-to-noise ratio and frequency analysis of continuous loop averaging deconvolution (CLAD) of overlapping evoked potentials,” J. Acoust. Soc. Am. 119, 429–438. Ozdamar, O., Delgado, R. E., Yavuz, E., Thombre, K., and Acikgoz, N. (2003a). “Deconvolution of auditory evoked potentials obtained at high stimulus rates,” in Proceedings of the First International IEEE EMBS Conference Neural Engineering (Capri, Italy), pp. 285–288. Ozdamar, O., Delgado, R. E., Yavuz, E., Thombre, K., and Anderson, M. (2003b). “Acquisition of ABRs at very high stimulation rate using CLAD (Continuous Loop Algorithm Deconvolution),” Assoc. Res. Otolaryngol. Abstr. 40, 44. Pratt, H., and Sohmer, H. (1976). “Intensity and rate functions of cochlear and brainstem evoked responses to click stimuli in man,” Arch. OtoRhino-Laryngol. 212, 85–92. Schimmel, H., Rapin, I., and Cohen, M. M. (1974). “Improving evoked response audiometry with special reference to the use of machine scoring,” Audiology 13, 33–65. Stapells, D. R., Picton, T. W., and Smith, A. D. (1982). “Normal hearing thresholds for clicks,” J. Acoust. Soc. Am. 72, 74–79. Stockard, J. J., Stockard, J. E., and Sharbrough, F. W. (1978). “Nonpathologic factors influencing brainstem auditory evoked potentials,” Am. J. EEG Technol. 18, 177–209. Stone, J. L., Calderon-Arnulphi, M., Watson, K. S., Patel, K., Mander, N. S., Suss, N., Fino, J., and Hughes, J. R. (2009). “Brainstem auditory evoked potentials: A review and modified studies in healthy subjects,” J. Clin. Neurophysiol. 26, 167–175. Thornton, A. R. D. (1993). “High rate otoacoustic emissions,” J. Acoust. Soc. Am. 94, 132–136. Thornton, A. R. D., Lineton, B., Baker, V. J., and Slaven, A. (2006). “Nonlinear properties of otoacoustic emissions in normal and impaired hearing,” Hear. Res. 219, 56–65. Thornton, A. R. D., and Slaven, A. (1993), “Auditory brainstem responses recorded at fast stimulation rates using maximum length sequences,” Br. J. Audiol. 27, 205–210. Valderrama, J. T., Alvarez, I., DeLaTorre, A., Segura, J. C., Sainz, M., and Vargas, J. L. (2011). “Educational approach of a BAER recording system based on experiential learning,” Technics Technol. Education Mgmt. 6, 876–889. Wang, T., Ozdamar, O., Bohorquez, J., Shen, Q., and Cheour, M. (2006). “Wiener filter deconvolution of overlapping evoked potentials,” J. Neurosci. Methods 158, 260–270. Weber, B. A., and Fletcher, G. L. (1980). “A computerized scoring procedure for auditory brainstem response audiometry,” Ear Hear. 1, 233–236. Webster, J. G., and Clark, J. W. (1995). Medical Instrumentation: Application and Design (John Wiley & Sons, New York), p. 814. Wong, P. K. H., and Bickford, R. G. (1980). “Brain stem auditory evoked potentials: The use of noise estimate,” Electroencephalogr. Clin. Neurophysiol. 50, 25–34. Yagi, T., and Kaga, K. (1979). “The effect of the click repetition rate on the latency of the auditory evoked brain stem response and its clinical use for a neurological diagnosis,” Arch. Oto-Rhino-Laryngol. 222, 91–96. Zollner, C., Karnahl, T., and Stange, G. (1976). “Input-output function and adaptation behaviors of the five early potentials registered with the earlobe-vertex pick-up,” Arch. Oto-Rhino-Laryngol. 212, 23–33.


3865


and infant hearing loss: Detection and intervention,” Pediatrics 103, 527–530. Eysholdt, U., and Schreiner, C. (1982). “Maximum length sequences: A fast method for measuring brain-stem-evoked responses,” Audiology 21, 242–250. Grandori, F., and Lutman, M. (1999). “The European Consensus Development Conference on Neonatal Hearing Screening (Milan, May 15–16, 1998),” Am. J. Audiology 8, 19–20. Hall, J. W. (2007). New Handbook of Auditory Evoked Responses (Allyn and Bacon, Boston, MA), p. 750. Hine, J. E., Ho, C.-T., Slaven, A., and Thornton, A. R. D. (2001). “Comparison of transient evoked otoacoustic emission thresholds recorded conventionally and using maximum length sequences,” Hear. Res. 156, 104–114. Jewett, D. L., Caplovitz, G., Baird, B., Trumpis, M., Olson, M. P., and Larson-Prior, L. J. (2004). “The use of QSD (q-sequence deconvolution) to recover superposed, transient evoked-responses,” Clin. Neurophysiol. 115, 2754–2775. Jewett, D. L., and Williston, J. S. (1971). “Auditory evoked far fields averaged from the scalp of humans,” Brain 94, 681–696. Jiang, Z. D., Brosi, D. M., Shao, X. M., and Wilkinson, A. R. (2000). “Maximum length sequence brainstem auditory evoked responses in term neonates who have perinatal hypoxia ischemia,” Pediatr. Res. 48, 639–645. Jiang, Z. D., Wu, Y. Y., and Wilkinson, A. R. (2009). “Age-related changes in BAER at different click rates from neonates to adults,” Acta Paediatr. 98, 1284–1287. Katz, J. (1994). Handbook of Clinical Audiology (Lippincott, Williams, Wilkins, Baltimore, MD), p. 1032. Kennedy, C. R., Kimm, L., Cafarelli-Dees, D., Evans, P. I. P., Hunter, M., Lenton, S., and Thornton, A. R. D. (1991). “Otoacoustic emissions and auditory brainstem responses in the newborn,” Arch. Dis. Child 66, 1124–1129. Kjaer, M. (1980). “Brain stem auditory and visual evoked potentials in multiple sclerosis,” Acta Neurol. Scand. 62, 14–19. Klein, A. J., and Teas, D. C. (1978). “Acoustically dependent latency shifts of BSER (wave V) in man,” J. Acoust. Soc. Am. 63, 1887–1895. Lasky, R. E. (1984). “A developmental study on the effect of stimulus rate on the auditory evoked brain-stem response,” Electroencephalogr. Clin. Neurophysiol. 59, 411–419. Lasky, R. E. (1997). “Rate and adaptation effects on the auditory evoked brainstem response in human newborns and adults,” Hear. Res. 111, 165–176. Lavoie, B. A., Barks, A., and Thornton, A. R. D. (2010). “Linear and nonlinear temporal interaction components of mid-latency auditory evoked potentials obtained with maximum length sequence stimulation,” Exp. Brain Res. 202, 231–237. Leung, S., Slaven, A., Thornton, A. R. D., and Brickley, G. J. (1998). “The use of high stimulus rate auditory brainstem responses in the estimation of hearing threshold,” Hear. Res. 123, 201–205. Lina-Granade, G., Collet, L., Morgon, A., and Salle, B. (1993). “Maturation and effect of stimulus rate on brainstem auditory evoked potentials,” Brain Dev. 15, 263–269. Mason, S. M., Su, A. P., and Hayes, R. A. (1977). “Simple online detector of auditory evoked cortical potentials,” Med. Biol. Eng. Comput. 15, 641–647. Millan, J., Ozdamar, O., and Bohorquez, J. (2006). “Acquisition and analysis of high rate deconvolved auditory evoked potentials during sleep,” in

Author's personal copy

Clinical Neurophysiology 125 (2014) 805–813

Contents lists available at ScienceDirect

Clinical Neurophysiology journal homepage: www.elsevier.com/locate/clinph

A study of adaptation mechanisms based on ABR recorded at high stimulation rate Joaquin T. Valderrama a,b,⇑, Angel de la Torre a, Isaac Alvarez a, Jose Carlos Segura a, A. Roger D. Thornton b, Manuel Sainz c,d, Jose Luis Vargas c a

Department of Signal Theory, Telematics and Communications, CITIC-UGR, University of Granada, C/ Periodista Daniel Saucedo Aranda s/n, 18071 Granada, Spain MRC Institute of Hearing Research, Southampton Outstation, Royal South Hants Hospital, Brintons Terrace, Mailpoint OAU, Southampton, Hampshire SO14 OYG, UK ENT Service, San Cecilio University Hospital, Av. Dr. Oloriz 16, 18002 Granada, Spain d Department of Surgery and its Specialties, University of Granada, Av. De Madrid 11, 18012 Granada, Spain b c

a r t i c l e

i n f o

Article history: Available online 13 October 2013 Keywords: Adaptation Auditory Brainstem Response (ABR) Evoked potentials Interstimulus interval (ISI) Randomized stimulation and averaging (RSA)

h i g h l i g h t s The fast and slow adaptation mechanisms are studied for the first time in humans through the sep-

arated responses methodology. Both fast and slow mechanisms of adaptation are present in all subjects, which is consistent with pre-

vious animal studies based on spike rate. The morphology of the ABR is not only influenced by the stimulation rate, but also by the distribution

of the jitter, and by the sequencing of stimuli.

a b s t r a c t Objective: This paper analyzes the fast and slow mechanisms of adaptation through a study of latencies and amplitudes on ABR recorded at high stimulation rates using the randomized stimulation and averaging (RSA) technique. Methods: The RSA technique allows a separate processing of auditory responses, and is used, in this study, to categorize responses according to the interstimulus interval (ISI) of their preceding stimulus. The fast and slow mechanisms of adaptation are analyzed by the separated responses methodology, whose underlying principles and mathematical basis are described in detail. Results: The morphology of the ABR is influenced by both fast and slow mechanisms of adaptation. These results are consistent with previous animal studies based on spike rate. Conclusions: Both fast and slow mechanisms of adaptation are present in all subjects. In addition, the distribution of the jitter and the sequencing of the stimuli may be critical parameters when obtaining reliable ABRs. Significance: The separated responses methodology enables for the first time the analysis of the fast and slow mechanisms of adaptation in ABR obtained at stimulation rates greater than 100 Hz. The non-invasive nature of this methodology is appropriate for its use in humans. Ó 2013 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved.

1. Introduction Adaptation of the auditory system is a decrease in response when a maintained stimulus or successive click stimuli are

⇑ Corresponding author at: Department of Signal Theory, Telematics and Communications, CITIC-UGR, University of Granada, C/ Periodista Daniel Saucedo Aranda s/n, 18071 Granada, Spain. Tel.: +34 958 240 840; fax: +34 958 240 831. E-mail addresses: [email protected] (J.T. Valderrama), [email protected] (A. de la Torre), [email protected] (I. Alvarez), [email protected] (J.C. Segura), [email protected] (A. Roger D. Thornton).

presented (Thornton and Coleman, 1975; Gillespie and Muller, 2009). Modeling of adaptation has unleashed controversy since Sorensen (1959) postulated that the decrease in the response could be associated with either a decrease in the number of active nerve fibers, or a decrease of their spike rate. Later, other authors suggested that the mechanisms of adaptation not only comprise the synapses of hair cells, but also the axonal transmission characteristics of the neurons that compose the auditory nerve (e.g., Chimento and Schreiner, 1991; Woo et al., 2009). Most of the authors agree on the combination of various types of mechanisms involved in the adaptation process whose effects are manifested in

1388-2457/$36.00 Ó 2013 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved. http://dx.doi.org/10.1016/j.clinph.2013.06.190


806

J.T. Valderrama et al. / Clinical Neurophysiology 125 (2014) 805–813

different time scales: fast adaptation occurs during the first few milliseconds following stimulus onset, whilst slow adaptation is about ten-fold slower (around 40–100 ms) (e.g., Eggermont, 1985; Yates et al., 1985; LeMasurier and Gillespie, 2005; Zhang et al., 2007). Slower effects of adaptation (up to several seconds after the beginning of stimulation) have also been analyzed by other authors (e.g., Javel, 1996). Although the relationship between these types of adaptation is still unclear, recent studies provide different models for fast and slow adaptation. Both physiological models conclude that adaptation reduces the opening of transduction channels in cochlear hair cells, limiting the flow of K+ and Ca2+ ions into the hair cell and therefore, reducing the probability of action potential generation (LeMasurier and Gillespie, 2005; Stauffer et al., 2005; Gillespie and Muller, 2009). A better understanding of the properties of adaptation in the auditory nerve may be useful for several clinical applications such as detecting certain peripheral lesions (e.g., acoustic neuroma) at an early stage or modeling the mechanotransduction process (conversion of a mechanical stimulus into an electrical response) (e.g., Don et al., 1977; Stockard et al., 1977; Yagi and Kaga, 1979). The most relevant methods proposed to examine the effects of adaptation are based on spike rate of the auditory nerve, on otoacoustic emissions (OAE), and on auditory brainstem response (ABR). The spike rate of hair cells can be measured in animals by microelectrodes inserted into the nerve fibers and in cochlear implant patients. Many studies have characterized adaptation as a decrease in spike rate when continuous stimulation is presented. These studies report different types of adaptation according to their temporal effect following stimulus onset: rapid adaptation (few milliseconds), short-term adaptation (tens of milliseconds), long-term adaptation (seconds), and very-long-term adaptation (minutes) (Westerman and Smith, 1984; Eggermont, 1985; Yates et al., 1985; Chimento and Schreiner, 1991; Javel, 1996). The recovery time from adaptation in auditory nerve fibers has also been defined in terms of spike rate (Young and Sachs, 1973; Yates et al., 1985). The non-invasive nature of the OAE and ABR methods makes them appropriate to study adaptation in humans. On one hand, the effects of adaptation in evoked otoacoustic emissions are manifested as a decrease in the amplitude of the response (e.g., Picton et al., 1993; Thornton, 1993; Lina-Granade and Collet, 1995; Hine et al., 2001). On the other hand, amplitudes of ABR waves decrease and latencies increase as a consequence of adaptation, especially in more central response components (e.g., Thornton and Coleman, 1975; Yagi and Kaga, 1979; Lasky, 1984; Jiang et al., 2009; Valderrama et al., 2012a). Conventional ABR recording technique consists of averaging several auditory responses whose corresponding stimuli are presented periodically. Many studies have used the conventional recording technique to analyze the effects of adaptation in ABR (e.g., Thornton and Coleman, 1975; Yagi and Kaga, 1979; Lasky, 1997, 1984; Polyakov and Pratt, 2003; Jiang et al., 2009). Some of these studies presented trains of clicks and recorded the transition from unadapted ABR to adapted ABR. The conventional technique has the limitation that the inter-stimulus interval (ISI) must be greater than the averaging window in order to avoid the contamination of the recording by the adjacent response (e.g., Kjaer, 1980). Thus, the conventional technique cannot be used to record ABR at rates higher than 100 Hz, considering a standard averaging window of 10 ms. However, the use of higher stimulation rates allows a more detailed study of adaptation since its effects increase with stimulus duration, stimulus level, and stimulation rate (e.g., Killian et al., 1994; Burkard et al., 1996a,b; Haenggeli et al., 1998; Clay and Brown, 2007; Zhang et al., 2007). It is not mathematically possible to recover the overlapped ABR signal when the stimulation sequence is periodic (conventional

stimulation) (e.g., Jewett et al., 2004). On this framework, different techniques have emerged to overcome the limitation imposed by the conventional technique. These techniques are able to obtain the superposed ABR signal using jittered stimuli (the jitter of a stimulation sequence measures the grade of dispersion of the ISI in contrast to a periodical presentation of stimuli where ISI would be constant). The most relevant techniques developed to record ABR at stimulation rates higher than 100 Hz are maximum length sequences (MLS) (Eysholdt and Schreiner, 1982), continuous loop averaged deconvolution (CLAD) (Delgado and Ozdamar, 2004; Ozdamar and Bohorquez, 2006), quasi-periodic sequence deconvolution (QSD) (Jewett et al., 2004), and randomized stimulation and averaging (RSA) (Valderrama et al., 2012a). The MLS technique has been widely used to explore the effects of adaptation in ABR recorded at high stimulation rates (e.g., Thornton and Slaven, 1993; Burkard et al., 1996a,b; Leung et al., 1998; Lavoie et al., 2008). The MLS, CLAD, and QSD techniques obtain the auditory response by averaging a number of blocks of responses corresponding to a predefined stimulation sequence, and then, deconvolving the response from the stimulation sequence by different procedures. The influence of the distribution of the jitter on the morphology of the auditory responses has not already been analyzed because the techniques based in deconvolution assume the premise that each click evokes the same response. The ABR recorded with RSA is obtained directly by averaging the responses after applying a digital blanking process which is useful for minimizing the effect of stimulation artifact. In comparison to CLAD, and QSD, the RSA technique allows a precise control of the jitter of the stimulation sequence, and a separate processing of auditory responses. This article presents a study of the fast and slow adaptation mechanisms based on ABR obtained with the RSA technique. Portions of this research were presented at the Adult Hearing Screening Congress, Cernobbio (Lake Como), Italy, June 7–9, 2012 (Valderrama et al., 2012b). The present study compares the amplitudes and latencies of waves III and V of the ABR obtained in different recording conditions. The stimulation sequences considered in this study are: (a) stimulation sequences with jitter distributions of long ISIs, (b) of short ISIs, and (c) of both long and short ISIs randomly distributed. The auditory responses corresponding to the long-and-short ISIs stimulation sequence were categorized according to the ISI of their preceding stimulus (long or short), and two ABR signals were obtained using these categories. If the morphology of the ABR-L and ABR-S signals (i.e. average of responses who’s preceding ISIs were long and short, respectively) were similar, that would suggest that the adaptation responds to slow mechanisms since the morphology of the ABR depends in a great extent on the stimulation rate of several preceding stimuli. On the other hand, if the morphology of ABR-L and ABR-S were similar to their corresponding ABR signals recorded with long and short ISI stimulation sequences, that would mean that the adaptation responds to fast mechanisms because the morphology of the response is strongly influenced by the ISI of the preceding stimulus. The results of this experiment show that most of the subjects analyzed in the study give results that lie in between both described situations, which suggests that both fast and slow mechanisms are involved in the adaptation process. The relevance of these findings is discussed in this article.

2. Methods 2.1. Subjects Eighteen subjects with no self-reported history of auditory dysfunction (normally hearing subjects), 4 females and 14 males, aged from 25 to 62 years (with a mean age of 34 years) participated in


807


this study. These subjects were chosen randomly from different social sectors from the University of Granada (e.g., students, professors, etc.). All subjects were volunteers and were informed about the experimental protocol and possible side effects of the test. A consent form was signed by the participants before the beginning of the recording session, which was carried out at the University of Granada (Granada, Spain) accordingly to The Code of Ethics of the World Medical Association (Declaration of Helsinki) for experiments involving humans. This recording procedure was approved by the Human Research Ethics Committee of the University of Granada and by the Clinical Research Ethics Committee of the San Cecilio University Hospital.

(A) ISIa−b

a

b

a

Time (ms)

The subjects were stimulated by clicks at an intensity of 70 dB above normal hearing level threshold (dBnHL). Monophasic clicks of 0.1 ms in condensation polarity were chosen as stimuli to evoke a synchronous firing of a large number of neurons, in particular those in the 1000–4000 Hz region (Hall, 2007; Thornton, 2007). The recording sessions were held in a room prepared to attenuate acoustical and electromagnetic interference. The subjects were seated in a comfortable position during the recording session in order to minimize the electromyogenic noise. Standard circumaural headphones (Pro-550, Ultrasone, Wielenbach, Germany) were used to present the stimuli to the subjects. The auditory evoked responses were recorded by three Ag/AgCl surface electrodes placed on the skin at the high forehead (active), ipsilateral mastoid (reference), and low forehead (ground). Interelectrode impedances were below 10 kO in all recordings. The signal recorded by the electrodes was amplified and band-pass filtered (100–3500 Hz). The band limits of the filters were chosen to maximize the detectability of all waves (Thornton, 2007). The synchronization of the biological

3 4

5

6

c

d

Time (ms)

signal with the stimuli was achieved through a synchronous recording of the EEG and the stimulation signal by a two-channel analogue-to-digital converter. Signals were sampled at 25 kHz and stored using 16 bits/sample. Data processing was carried out by algorithms implemented in MATLAB (The Mathworks, Inc., Natick, MA). The recorded EEG was digitally filtered using a sixth order bandpass Butterworth filter (150–3000 Hz). A full description of the ABR recording system can be found in Valderrama et al. (2011). 2.3. ABR obtained with RSA The recording of ABR at high stimulation rates using the RSA technique is appropriate to analyze the effects of adaptation. The ABR signal is obtained in RSA by averaging auditory responses corresponding to stimuli whose ISI varies randomly according to a

(A) ISI2−5 / 21−24 Sequence 2

b

Fig. 1. Distribution of the jitter for the two types of stimulation sequences used in this study. (A) Histogram of the interstimulus interval (ISI) for an ISIa–b stimulation sequence: the ISI varies uniformly random within the interval [a, b] ms. (B) Histogram of the ISI for an ISIa–b/c–d stimulation sequence: the ISI varies uniformly random within the intervals [a, b] and [c, d] ms.

2.2. EEG recording

1

(B) ISIa−b/c−d

(B) ISI2−5 / 21−24 Histogram

7 8

Stimulation seq. 1

2

3

4

6

1

2

3+4

5+6

20

40

5

7

Long ISI contrib. 8

Short ISI contrib. 7+8

Recorded signal

0 0

60

80

100

5

10

120

15 20 ISI (ms)

25

30

Time (ms)

Split & Average

(D) ISI21−24 ABR (split) Amplitude (µV)

Amplitude (µV)

(C) ISI2−5 ABR (split) 0.2 0.1 0 −0.1 −0.2 0

2

4

6

Time (ms)

8

10

0.2 0.1 0 −0.1 −0.2 0

2

4

6

8

10

Time (ms)

Fig. 2. Scheme of the process of separated responses. (A) Frame of an ISI2–5/21–24 stimulation sequence. The auditory response contribution without noise from each stimulus is categorized according to their preceding ISI. The stimuli and their associated auditory responses are numerated. The ‘‘Long ISI contrib.’’ and ‘‘Short ISI contrib.’’ signals shows the auditory responses corresponding to the stimuli whose preceding ISI belong to the intervals [21,24] and [2,5] ms, respectively. The ‘‘Recorded signal’’ shows the sum of both long and short ISI ABR contributions. (B) Histogram of the interstimulus interval for an ISI2–5/21–24 stimulation sequence. (C) ABR obtained with the auditory responses whose preceding ISI belong to the interval [2,5] ms. (D) ABR obtained with the auditory responses whose preceding ISI belong to the interval [21,24] ms.


808


Subject 1 III

Subject 2

V

Subject 3

V

III

V

III

21−24 (r) 21−24 (s) 2−5 (s)

0.15 µV

2−5 (r) 0

2

4

6

8

10

0

2

4

Subject 4 III

6

8

10

0

2

4

Subject 5 III

V

6

8

10

Subject 6 III

V

V

21−24 (r) 21−24 (s) 2−5 (s)

0.15 µV

2−5 (r) 0

2

4

6

8

10

0

2

4

Subject 7 III

6

8

10

0

2

4

Subject 8 III

V

6

8

10

Subject 9

V

V

III

21−24 (r) 21−24 (s) 2−5 (s)

0.15 µV

2−5 (r) 0

2

4

6

8

10

0

2

4

Subject 10

8

10

0

2

Subject 11

V

III

6

6

8

10

Subject 12

V

III

4

III

V

21−24 (r) 21−24 (s) 2−5 (s)

0.15 µV

2−5 (r) 0

2

4

6

8

10

0

2

4

Subject 13

8

10

0

2

Subject 14

V

III

6

III

4

6

8

10

Subject 15

V

V

III

21−24 (r) 21−24 (s) 2−5 (s)

0.15 µV

2−5 (r) 0

2

4

6

8

10

0

2

4

Subject 16 III

6

8

10

0

2

Subject 17 III

V

4

6

8

10

Subject 18

V

III

V

21−24 (r) 21−24 (s) 2−5 (s)

0.15 µV

2−5 (r) 0

2

4 6 Time (ms)

8

10

0

2

4 6 Time (ms)

8

10

0

2

4 6 Time (ms)

8

10

Fig. 3. ABR signals from 18 subjects obtained in the following recording conditions. The ‘21–24 (r)’ and ‘2–5 (r)’ signals represent the recorded ABRs obtained using the randomized stimulation and averaging (RSA) technique with the stimulation sequences ISI21–24 and ISI2–5, respectively. The ‘21–24 (s)’ and ‘2–5 (s)’ represent the separated ABRs obtained with the separated responses methodology on the stimulation sequence ISI2–5/21–24. Waves III and V are identified in all recordings.


809


Table 1 Interval and mean (standard deviation in parentheses) of the latencies (L) and amplitudes (A) of the waves III and V from the ABR signals presented in the Fig. 3. Latencies and amplitudes are measured in milliseconds and microvolts, respectively. This table shows that the averaged amplitudes and latencies of the separated ‘21–24 (s)’ and ‘2–5 (s)’ ABR signals are in between their corresponding ABR recorded signal and their opposite separated ABR signal. 21–24 (r)

LIII (ms) LV (ms) AIII (lV) AV (lV)

21–24 (s)

2–5 (s)

2–5 (r)

Interval

Mean (S.D.)

Interval

Mean (S.D.)

Interval

Mean (S.D.)

Interval

Mean (S.D.)

[3.24 [5.40 [0.13 [0.12

3.60 5.79 0.23 0.19

[3.36 [5.56 [0.10 [0.12

3.68 5.93 0.18 0.17

[3.32 [5.64 [0.05 [0.05

3.77 6.11 0.11 0.10

[3.48 [6.12 [0.04 [0.05

3.93 6.72 0.08 0.08

3.86] 6.08] 0.32] 0.36]

(0.15) (0.19) (0.05) (0.06)

3.94] 6.20] 0.28] 0.26]

(0.15) (0.19) (0.05) (0.04)

3.96] 6.36] 0.15] 0.16]

(0.17) (0.21) (0.03) (0.03)

4.10] 7.12] 0.15] 0.17]

(0.16) (0.28) (0.03) (0.03)

Table 2 Interval, mean (standard deviation in parentheses) and p-value of the differences of latencies (L) and ratio of amplitudes (A) between pairs of ABR from each subject obtained in different conditions in a group of 18 subjects. This analysis suggests that all ABRs compared in this study are statistically different (p-value < 0.05) in terms of amplitudes and latencies, in exception for AV in ‘21–24 (s)/21–24 (r)’. 21–24 (s) – 21–24(r) Interval

Mean (S.D.)

2–5 (s) – 21–24(s) p-value

Interval

LIII (a) – LIII (b) (ms) LV (a) – LV (b) (ms)

[ 0.08 0.20] 0.08 (0.06) [ 0.02 0.30] 0.15 (0.08) 21–24 (s)/21–24(r)

4 10 7 10

5

AIII (a)/AIII (b) AV (a)/AV (b)

[0.60 1.07] [0.50 1.33]

7 10 0.407

4

0.82 (0.14) 0.95 (0.23)

7

Mean (S.D.)

2–5 (r) – 2–5 (s) p-value

[ 0.04 0.34] 0.09 (0.09) [0.00 0.30] 0.18 (0.08) 2–5 (s)/21–24(s)

8 10 2 10

4

[0.36 0.93] [0.35 0.86]

2 10 210 4

4

predefined probability distribution. The RSA technique includes a digital blanking process and non-uniform averaging which considers as null values those samples contaminated by the stimulation artifact (0.2 ms before and 0.8 ms after each stimulus). These null values are not considered in the averaging process. A basic artifact rejection technique was used to improve the quality of the recordings: auditory responses whose amplitude exceeded the range ±10 lV were not considered in the averaging process. The RSA technique is described in detail in Valderrama et al. (2012a). The RSA technique allows a precise control of the jitter in the process of stimulation sequences generation. The stimulation sequences used in this study present two types of jitter distributions. The distribution of the jitter for each type of stimulation sequence is presented in Fig. 1. This study involves ISIa–b stimulation sequences, whose ISI varies randomly with an uniform distribution within an interval ‘a’ to ‘b’ ([a, b]) ms (Fig. 1A); and ISIa–b/c–d stimulation sequences, whose ISI varies with a uniform random distribution between the intervals ‘a’ to ‘b’ ([a, b]) and ‘c’ to ‘d’ ([c, d]) ms (Fig. 1B). 2.4. Separated responses The separated responses methodology is based in a separate processing of auditory responses, which can be performed using the RSA technique. Fig. 2 outlines the process of separating the responses. Fig. 2A shows a frame from an ISI2–5/21–24 stimulation sequence and their associated auditory responses without noise. The ISI of this stimulation sequence varies with a uniform random distribution between the intervals [2,5] and [21,24] ms, as shown by its histogram in Fig. 2B. The auditory responses can be categorized according to their preceding ISI. The auditory responses whose preceding ISI belong to the interval [21,24] ms (associated to stimuli 1, 2, 3, 5, 7) are shown as ‘‘long ISI contribution’’, and those whose preceding ISI belong to the interval [2,5] ms (associated to stimuli 4, 6, and 8) are shown as ‘‘short ISI contribution’’. The ‘‘recorded signal’’ in Fig. 2A shows the sum of both long and short ISI contributions. Fig. 2C and D show the ABR obtained using the RSA technique with the auditory responses that belong to each interval.

0.60 (0.17) 0.60 (0.13)

8

Interval

Mean (S.D.)

p-value

[ 0.02 0.56] [0.30 0.84] 2–5 (r)/2–5 (s)

0.17 (0.13) 0.61 (0.18)

5 10 9 10

[0.45 1.50] [0.54 1.21]

0.82 (0.29) 0.81 (0.20)

0.020 0.004

5 11

2.5. Description of the experiments The following EEGs were recorded from each subject: 5000 auditory responses corresponding to an ISI21–24 stimulation sequence, 10,000 auditory responses corresponding to an ISI2–5/ 21–24 stimulation sequence, and 20,000 auditory responses corresponding to an ISI2–5 stimulation sequence. The auditory responses were recorded, stored and processed offline. The number of recorded responses increases at higher stimulation rates because the quality of the ABR degrades as stimulation rate increases as a consequence of adaptation (e.g., Don et al., 1977; Valderrama et al., 2012a), and therefore, more auditory responses are needed in order to obtain ABR signals of similar quality. From the EEG corresponding to an ISI2–5/21–24 stimulation sequence, two ABR signals were obtained after the separated responses procedure described in Section 2.4. Thus, these two separated ABR signals were obtained with approximately 5000 auditory responses. The amplitudes and latencies of the waves III and V were measured as a difference in milliseconds between the top of the peaks and the stimulus onset for latencies, and the amplitudes as the difference in microvolts between the top of the peak and the following trough (Thornton, 2007; Hall, 2007). The mean and standard deviation of the amplitudes and latencies were calculated among the 18 subjects. The separated ABR responses and the recorded ABR responses were compared in terms of latencies by a matched paired t-test and in terms of amplitudes by a matched paired Wilcoxon signed rank test. Two hypotheses are considered in this study: (1) the recorded ISI21–24 ABR is similar to the separated ISI21–24 ABR and the recorded ISI2–5 ABR is similar to the separated ISI2–5 ABR (the two separated ABRs are different); and (2) both separated ISI21–24 and ISI2–5 ABRs are similar. On one hand, hypothesis 1 would indicate that the auditory system adapts according to fast mechanisms since the morphology of the separated ABR would be very much influenced by the ISI of the preceding stimulus. On the other hand, hypothesis 2 would suggest that adaptation is a slow process which is mostly influenced by the stimulation rate of several preceding stimuli (the influence of the preceding stimulus is not determinative). This paper also includes a study that analyzes the effect of the slow mechanisms of adaptation on the morphology of the


810

J.T. Valderrama et al. / Clinical Neurophysiology 125 (2014) 805–813 III

V

Subject 1

V

III

V

III

Subject 2

Subject 3

[2−5] & [21−24] 0.15 µV [2−5/21−24] 0

2

4 III

6 V

8

10

0

2

4

6

III

Subject 4

V

8

10

0

2

4 V

III

Subject 5

6

8

10

Subject 6

[2−5] & [21−24] 0.15 µV [2−5/21−24] 0

2

4

6

III

8

10

0

2

6 V

III

Subject 7

[2−5] & [21−24]

4

8

10

0

2

4 III

Subject 8

6

8

V

Subject 9

10

V

0.15 µV

[2−5/21−24] 0

2

4

6

III

V

8

10

0

2

4 III

Subject 10

6 V

8

10

0

2

4

V

III

Subject 11

6

8

10

Subject 12

[2−5] & [21−24] 0.15 µV [2−5/21−24] 0

2

4

6 V

III

8

10

0

2

4

6

III

Subject 13

V

8

10

0

2

4 III

Subject 14

6 V

8

10

Subject 15

[2−5] & [21−24] 0.15 µV [2−5/21−24] 0

2

4 III

6 V

8

10

0

2

4 III

Subject 16

6 V

8

10

0

2

Subject 17

4 III

6 V

8

10

Subject 18

[2−5] & [21−24] 0.15 µV [2−5/21−24] 0

2

4 6 Time (ms)

8

10

0

2

4 6 Time (ms)

8

10

0

2

4 6 Time (ms)

8

10

Fig. 4. ABR signals from 18 subjects obtained using 10,000 auditory responses from stimulation signals of equal distributions of the jitter but different order of presentation of stimuli: whilst the ISI in the [2–5/21–24] stimulation sequence vary uniformly random between the ranges [2,5] and [21,24] ms all along the stimulation sequence, the ISI in the [2–5] and [21–24] stimulation sequence vary uniformly random between the range [21–24] ms during the first 5000 stimuli, and between the range [2,5] ms during the last 5000 stimuli. Waves III and V are labeled in the figure.

ABR. ABRs from 18 subjects obtained with 10,000 stimuli from an ISI2–5/21–24 stimulation signal (ABR [2–5/21–24]) were compared to ABRs obtained by averaging 5000 auditory responses from an ISI21–24 stimulation sequence and 5000 auditory responses from an ISI2–5 stimulation sequence (ABR [2–5]&[21–24]). These two ABRs are obtained with stimulation sequences of the same distribution of the jitter, but a different sequencing of stimuli. On the ABR [2–5/21–24], the ISI of the stimuli varies uniformly random between the ranges [2,5] and [21,24] ms all along the stimulation sequence; whilst on the ABR [2–5]&[21–24], the ISI varies uniformly random between [21,24] ms during the first 5000 stimuli and between [2,5] ms during the last 5000 stimuli. Considering that the fast mechanisms of adaptation are manifested within the few milliseconds following stimulus onset, these two ABR signals are influenced in the same manner by the fast mechanisms of adaptation since both ABRs involve 5000 auditory responses whose preceding ISI belong to the interval [2,5] ms and 5000 responses whose preceding ISI belong to the interval [21,24] ms. The two ABR signals of this experiment will be different according to the

effects of the slow mechanisms of adaptation. The slow mechanisms of adaptation are manifested, on one hand, during 10,000 responses at an averaged ISI of 13 ms on the ABR [2–5/21–24]; and on the other hand, at an averaged ISI of 22.5 ms during the first 5000 responses and at an averaged ISI of 3.5 ms during the last 5000 responses on the ABR [2–5]&[21–24]. A statistical difference among the two ABR signals obtained with this experimental protocol could be used to detect the influence of the slow mechanisms of adaptation on the ABR. 3. Results Fig. 3 shows ABR signals obtained from the group of 18 subjects in the previously described recording conditions. The recorded ABR signals corresponding to the ISI21–24 and ISI2–5 stimulation sequences are represented by ‘21–24 (r)’ and ‘2–5 (r)’, respectively; and the separated ABR signals are represented by ‘21–24 (s)’ and ‘2–5 (s)’. The waves III and V are labeled in the figure and were identified in all subjects. Despite the differences in the morphology among ABR from different subjects, this figure shows that most of


811


the subjects present a similar pattern, which is analyzed in Tables 1 and 2. Table 1 shows the mean and standard deviation of the latencies and amplitudes of waves III and V in a group of 18 subjects. The amplitudes and latencies measured on the recordings ‘21–24 (r)’ and ‘2–5 (r)’ are consistent with previous literature (e.g., Yagi and Kaga, 1979; Lasky, 1984; Lina-Granade et al., 1993; Leung et al., 1998; Jiang et al., 2009; Valderrama et al., 2012a,b). This table indicates that both amplitudes and latencies are influenced by the stimulation rate: amplitudes decrease and latencies increase as stimulation rate increases as a consequence of adaptation. The effects of adaptation in latencies are more remarkable in wave V than in wave III, since the stimulation rate influences in a greater extent those components generated in a more central site (e.g., Pratt and Sohmer, 1976; Yagi and Kaga, 1979; Jiang et al., 2009; Valderrama et al., 2012a). This table shows that, on average, the recorded ‘21–24 (r)’ ABR signals present greater amplitudes and lower latencies than the ‘2–5 (r)’ signals in both waves; that the separated ‘21–24 (s)’ ABR signal presents amplitudes and latencies in between the ‘2–5 (s)’ and the ‘21–24 (r)’ ABR signals; and that the ‘2–5 (s)’ ABR signal presents amplitudes and latencies in between the ‘21–24 (s)’ and the ‘2–5 (r)’ ABR signals. Table 2 compares the amplitudes and latencies of waves III and V from pairs of ABRs from each subject and analyzes whether or not their differences are statistically significant. The latencies and amplitudes are analyzed in this table in terms of differences and ratio, respectively. Table 2 shows the mean and standard deviation of the differences of latencies and ratio of amplitudes between the following pairs of ABRs: ‘21–24 (r)’ vs ‘21–24 (s)’, ‘21–24 (s)’ vs ‘2– 5 (s)’, and ‘2–5(r)’ vs ‘2–5 (s)’. The p-value shown in the table indicates the probability of obtaining those results by chance, considering as reference differences of latencies equal to zero and ratio of amplitudes equal to one. The large standard deviation of these parameters points out a large variability among subjects. This table also shows that there are statistically significant differences (p-value < 0.05) between (a) both separated ‘21–24 (s)’ and ‘2–5 (s)’ ABR signals in terms of amplitudes and latencies, and (b) between each separated ABR signals and its corresponding recorded ABR signals. The morphology of the recordings ‘21–24 (s)’ and ‘21–24 (r)’ may be assumed to be different despite their ratio of AV does not show statistically significant differences (p-value > 0.05), since the rest of parameters (AIII, LIII, and LV) are statistically different. Fig. 4 shows ABR signals from 18 subjects corresponding to a stimulation sequence in which the ISI varies uniformly random between the ranges [2,5] and [21,24] ms (shown as ‘2–5/21–24’ in the figure); and corresponding to a stimulation sequence in which the ISI vary between the interval [21,24] ms during the first 5000 stimuli, and between the interval [2,5] ms during the last 5000 stimuli (shown as ‘[2–5]&[21–24]’ in the figure). The mean and standard deviation of the amplitudes and latencies of the waves III and V in these recordings are shown in Table 3. Waves III and V could be identified in all signals, in exception for the wave V in subject 7 in the [2–5]&[21–24] ABR signal. The two ABR recordings from each subject were compared with a matched paired t-test for

Table 3 Interval and mean (standard deviation in parentheses) of the latencies (L) and amplitudes (A) of waves III and V from the ABR signals presented in Fig. 4. [2–5/21–24]

LIII (ms) LV (ms) AIII (lV) AV (lV)

[2–5] & [21–24]

Interval

Mean (S.D.)

Interval

Mean (S.D.)

[3.36 [5.58 [0.08 [0.06

3.70 5.96 0.14 0.13

[3.28 [5.44 [0.06 [0.04

3.63 5.80 0.12 0.09

3.92] 6.20] 0.20] 0.21]

(0.15) (0.19) (0.04) (0.04)

3.94] 6.04] 0.18] 0.17]

(0.17) (0.18) (0.03) (0.03)

Table 4 Interval, mean (standard deviation in parentheses) and p-value of the differences of latencies (L) and ratio of amplitudes (A) between pairs of ABR from each subject obtained in different recording conditions. This table remarks that there are statistically significant differences between the [2–5/21–24] and the [2–5]&[21–24] ABR signals (p-value < 0.05). [2–5/21–24] – [2–5]&[21–24]

LIII (a) – LIII (b) (ms) LV (a) – LV (b) (ms)

Interval

Mean (S.D.)

p-value

[ 0.08 0.24] [0.04 0.26]

0.07 (0.07) 0.16 (0.05)

610 4 910 10

[2–5/21–24]/[2–5]&[21–24] Mean (S.D.) AIII (a)/AIII (b) AV (a)/AV (b)

[0.75 2.00] [0.60 3.00]

1.24 (0.29) 1.57 (0.69)

p-value 0.003 0.003

differences of latencies and with a matched Wilcoxon signed rank test for ratio of amplitudes. The analysis for waves III and V were made with 18 and 17 subjects, respectively. The results of this study are presented in Table 4. This table shows that there are statistically significant differences between the ‘[2–5/21–24]’ and the ‘[2–5]&[21–24]’ ABR signals, which confirms the influence of the slow mechanisms of adaptation on the morphology of the auditory response. 4. Discussion This article presents a study of the fast and slow mechanisms of adaptation based on ABR signals obtained at high stimulation rates using the RSA technique. The recorded ‘21–24 (r)’ and ‘2–5 (r)’ ABR signals were obtained using directly the RSA technique with auditory responses whose ISI varied randomly within the range [21,24] and [2,5] ms, respectively. The separated ‘21–24 (s)’ and ‘2–5 (s)’ ABR signals were obtained using the separated responses methodology with the EEG corresponding to the ISI2–5/21–24 stimulation sequence, which allows the retrieval of auditory responses whose preceding ISI belong to the interval [21,24] ms or to the interval [2,5] ms. The comparison of ABR signals was carried out by an analysis of the differences in latencies and ratio of amplitudes. If the separated ABR signals were similar to their corresponding recorded ABR signals (both separated ABRs were different), the fast mechanisms of adaptation would prevail over the slow mechanisms since the morphology of the response would be influenced in a greater extent by the ISI of the preceding stimulus. On the other hand, if the separated ABR signals were different to their corresponding recorded ABR signals and both separated ABRs were similar, the slow mechanisms of adaptation would have prevailed over the fast mechanisms because the morphology of the response would be mainly determined by the averaged stimulation rate of several preceding stimuli (but not by the ISI of the preceding stimulus). The results of this study indicate that most of the subjects present a situation in between both hypotheses, which suggests that both fast and slow mechanisms of adaptation influence the morphology of the auditory response. There exists a great variability among subjects (Fig. 3). For instance, the separated ABR signals in subjects 10 and 17 present high differences in amplitudes but small differences in latencies; subjects 1 and 5 present high differences in both amplitudes and latencies; and subjects like 15 and 16 show small differences in amplitudes but high differences in latencies. On average, the latencies and amplitudes of the main waves in the ‘21–24 (s)’ and ‘2–5 (s)’ ABR signals are, respectively, in between the ‘2–5 (s)’ and the ‘21–24 (r)’ ABR signals on one hand, and between the ‘21–24 (s)’ and the ‘2–5 (r)’ on the other hand (see Table 1). The results presented in Table 2 show that the two separated ABR signals are statistically different, and that there


812


are statistically significant differences between the separated ABR responses and their corresponding recorded ABR responses (see Table 2). These findings indicate that the morphology of the separated ABR is influenced both by the ISI of the preceding stimulus and by the averaged stimulation rate of several preceding stimuli, which suggests that both fast and slow mechanisms are involved in the adaptation process. This paper also includes an experimental protocol to detect the influence of the slow mechanisms of adaptation on the morphology of ABR. The results presented in Fig. 4 and Tables 3 and 4 show statistically significant differences between ABRs obtained with long and short ISI clicks randomly presented all along the stimulation sequence (averaged ISI of 13 ms) and ABRs obtained with long ISI clicks in the beginning (averaged ISI of 22.5 ms) and short ISI clicks in the end (averaged ISI of 3.5 ms). These results confirm the existence of slow mechanisms of adaptation in ABR. In addition, these results indicate that the morphology of the ABR is not only influenced by the average stimulation rate, but also by the distribution of the jitter and the sequencing of the stimuli. The results presented in this paper are consistent with previous studies, in which the fast and slow mechanisms of adaptation are characterized in animals in terms of spike rate (e.g., Westerman and Smith, 1984; Eggermont, 1985; Yates et al., 1985; Javel, 1996). The fast mechanisms of adaptation analyzed in this study are manifested during the first few milliseconds following stimulus onset and may be related to the rapid adaptation described in Westerman and Smith (1984) and in Yates et al. (1985). Although the time constant for the slow mechanisms of adaptation is not determined in this paper, the results presented in Fig. 4 and Tables 3 and 4 indicate that the time constant for the slow mechanisms of adaptation might be greater than 20 ms, otherwise the effects of slow adaptation would not have been observed in that experiment. The slow mechanisms of adaptation observed in these experiments may be related to the short-term adaptation defined in Westerman and Smith (1984) and to the long-term adaptation described in Javel (1996), whose time constant varies from several tens of milliseconds to a few seconds. The non-invasive nature of the process of ABR recording is appropriate to study the effects of adaptation in humans. Traditionally, the adaptation of the hearing system was analyzed by presenting to the subject trains of stimuli of a fixed ISI, and comparing the morphology of the ABRs corresponding to each position in the train (e.g., Thornton and Coleman, 1975; Lasky, 1997; Polyakov and Pratt, 2003). This methodology presents the limitation that the ISI must be greater than the averaging window. Thus, the adaptation cannot be studied using this methodology at rates greater than 100 Hz. Other techniques like MLS, CLAD, or QSD allow the recording of ABR at very high stimulation rates (Eysholdt and Schreiner, 1982; Delgado and Ozdamar, 2004; Jewett et al., 2004; Ozdamar and Bohorquez, 2006). These techniques obtain the ABR signal through jittered stimuli and different deconvolution processes, which require the processing of sets of responses, and therefore, limit the study of the fast and slow mechanisms of adaptation since they assume that each click evokes the same response. The separated responses methodology performed with RSA allows for the first time a separate processing of auditory responses at stimulation rates greater than 100 Hz, which can be used to study the fast and slow effects of adaptation. The flexible control of the distribution of the jitter, the design of the sequencing of stimuli, and a separate processing of auditory responses are advantages of the RSA methodology that may be of interest in the design of certain experiments in audiology. Despite that both fast and slow mechanisms of adaptation studied in this article seem to be related to changes in the auditory mechanotransduction, the origin of such mechanisms may be analyzed separately. It is generally accepted a time boundary, at

approximately 50 ms, to separate components affected by attention (endogenous components, latencies >50 ms) and those that are not (exogenous components, latencies 0.9 in all measures), (b) that latencies estimated by FPP are accurate since the linear regression curves are close to the curves FPP = MAN (dotted line), and (c) that a slight bias exists between the amplitudes measured manually and automatically by FPP, possibly as a consequence of local noise, which systematically provokes an overestimation of amplitudes by the manual method. Fig. 4 shows examples of ABR signals used in this experiment from 5 subjects at different stimulation rates. The parametric peaks adjusted to the waves III

and V are highlighted on this figure. In addition, this figure includes the SNR associated with each peak evaluated automatically by the FPP method. Table 2 presents the mean and standard deviation of the latencies, amplitudes, widths, and SNRs measured automatically by the FPP method on the waves III and V. This table shows the tendency of the parameters as stimulation rate increases: latencies increase, the interpeak latency between waves III and V increases because the shift of wave V is greater than in wave III, the amplitudes of both waves decrease, the widths increase in both waves, possibly as a consequence of neural desynchronization [47], and the SNRs of both waves tend to decrease due to the lower amplitude of the waves. Table 3 presents the mean and standard deviation of the latencies and amplitudes measured manually on waves III and V. The analysis of Tables 2 and 3 shows that, on average, there are similarities between the values measured manually and automatically by the FPP method on the latencies of waves III and V, and on the amplitude of wave III. Regarding the amplitude of wave V, there is a systematic difference of a few tens of nanovolts on the values measured manually and automatically by the FPP method. This difference might arise because the trough that follows wave V does not fit perfectly the template. Nonetheless, the values of the latencies, amplitudes, and widths shown in both Tables 2 and 3 are consistent with those reported in previous studies [4,48–52].

3.3.

Experiment 2

3.3.1.

Subjects and methods

In this second experiment, the performance of the automatic quality assessment based on the FPP method is compared to the automatic quality evaluation techniques based on the correlation coefficient (r), the FSP , and the cross correlation with a predefined template method (Cross Corr). The ABR signals used in this test consisted of 500 recordings from 10 normal hearing subjects (6 males and 4 females; aged between 21 and 37 years). Each recording was obtained with auditory stimuli periodically presented at a rate of 30 Hz, at a different number of averaged sweeps (100, 300, 900, 1800, and 9500). From these 500 recordings, 40 recordings were obtained without auditory stimulation, so no ABR could be detected. The correlation coefficient (r) analysis was performed on the interval [1,10] ms to minimize the effect of the recorded artifacts synchronized with the stimulus. The single point (SP) chosen for the implementation of the FSP method was the sample 100 (corresponding to the 4th ms of the averaging window, considering fs = 25 kHz). The template waveform used on the Cross Corr method was built from ABR signals recorded on 30 normal hearing subjects (17 males and 13 females; aged between 17 and 34 years) in the same recording conditions as the test signals, using 2000 averaged sweeps. The template waveform used in the Cross Corr method is available as supplementary material (section B). These subjects were different from those analyzed to obtain the ABR signals used for test. Each ABR signal used to build the template waveform was normalized in amplitude according to its RMS value, cosined-tapered with a band pass window of [1,8] ms, and scaled in amplitude producing an RMS value equal to the mean of the RMS values of the original recordings. The mean of these signals produced the

Author's personal copy 267

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 4 ( 2 0 1 4 ) 262–275

RAW data 9

0.55 2

LIII MAN (R =0.49)

A MAN (R =0.20)

7 6.5 6 5.5 Wave III 5

0.45

Amplitude of wave V (µV)

Wave V

V

0.4 0.35 0.3 0.25 0.2 0.15

4.5

0.4 0.35 0.3 0.25 0.2 0.15

4

0.1

0.1

3.5

0.05

0.05

40

70

2

A MAN (R =0.24) V

0.45

Amplitude of wave III (µV)

V

L MAN (R2=0.60)

7.5

V

0.5

2

III

2

L FPP (R =0.71)

A FPP (R =0.34)

III

0.5

2

8

2

A FPP (R =0.27)

III

8.5

Latencies of waves III and V (ms)

0.55

L FPP (R2=0.59)

100 130 160 190 220 250

40

70

Stimulation rate (Hz)

100 130 160 190 220 250

40

70


100 130 160 190 220 250


Normalized data 9

0.55

0.55

2

III

L FPP (R2=0.96)

0.45

2

L MAN (R =0.85) V

Wave V

7 6.5 6 5.5 Wave III 5

V

0.45

0.4 0.35 0.3 0.25 0.2

0.4 0.35 0.3 0.25 0.2

4.5

0.15

0.15

4

0.1

0.1

3.5

40

70

A MAN (R2=0.48)

III

V

7.5

AV FPP (R2=0.62)

0.5

A MAN (R2=0.62)

Amplitude of wave V (µV)

8

LIII MAN (R2=0.65)

0.5

Amplitude of wave III (µV)

Latencies of waves III and V (ms)

AIII FPP (R2=0.66)

L FPP (R =0.80)

8.5

100 130 160 190 220 250

0.05

40


70

100 130 160 190 220 250

0.05


40

70

100 130 160 190 220 250


Fig. 2 – Latencies (L) and amplitudes (A) of waves III and V measured manually (MAN) and automatically by the FPP method in a set of 320 ABR signals obtained from eight normal hearing subjects at different stimulation rates. Normalized data in terms of the mean value are also presented on this figure to decrease the intersubject variability. The coefficients of determination (R2 ) obtained in this study on each parameter suggest that the model of amplitudes and latencies is better described with the FPP method.

template waveform used in the Cross Corr method. Further details of the implementation of the methods based on the correlation coefficient (r), on the FSP , and on the Cross Corr can be found, respectively, in [35,36,40]. The results obtained with the automatic methods were compared to a subjective evaluation provided by 5 experts. Each expert had at least three years of expertise in the analysis of ABR signals. The experts were asked to rate the quality

of a number of ABR signals according to the following criteria: Q = 0, no ABR is observed (no auditory response); Q = 1, wave V can be hardly detected (highly noisy ABR); Q = 2, wave V can be detected but the rest of waves are unclear (noisy ABR); Q = 3, waves III and V can be clearly detected (ABR slightly noisy); Q = 4, waves I, III, and V can be detected (good quality ABR); and Q = 5, all components of the ABR can be easily detected (excellent quality ABR). A computer application was programmed

Table 2 – Mean (and standard deviation in parentheses) of the latencies (L), amplitudes (A), widths (W), and SNRs of waves III and V measured automatically by the FPP on a set of 320 ABR signals obtained from 8 normal hearing subjects at different stimulation rates. Latencies and widths are measured in ms, amplitudes in ␮V, and SNR in dB. Stimulation rate (Hz) 45 55 72 83 100 125 167 250

LIII 3.74 (0.13) 3.79 (0.10) 3.86 (0.13) 3.91 (0.12) 3.92 (0.15) 4.01 (0.17) 4.18 (0.15) 4.33 (0.25)

LV

LV –LIII

AIII

AV

WIII

WV

SNRIII

5.71 (0.20) 5.80 (0.20) 5.91 (0.18) 5.98 (0.19) 6.09 (0.22) 6.21 (0.20) 6.41 (0.26) 6.77 (0.25)

1.97 (0.15) 1.99 (0.15) 2.04 (0.15) 2.06 (0.14) 2.12 (0.15) 2.21 (0.16) 2.19 (0.23) 2.42 (0.15)

0.25 (0.08) 0.23 (0.07) 0.21 (0.07) 0.20 (0.08) 0.17 (0.07) 0.14 (0.06) 0.15 (0.07) 0.11 (0.05)

0.26 (0.07) 0.25 (0.08) 0.22 (0.07) 0.21 (0.07) 0.19 (0.06) 0.18 (0.05) 0.15 (0.04) 0.13 (0.04)

0.37 (0.05) 0.38 (0.04) 0.37 (0.05) 0.38 (0.05) 0.38 (0.06) 0.40 (0.08) 0.52 (0.14) 0.54 (0.12)

0.46 (0.05) 0.47 (0.06) 0.49 (0.06) 0.53 (0.08) 0.50 (0.07) 0.50 (0.07) 0.52 (0.08) 0.58 (0.10)

9.19 (2.83) 8.30 (3.37) 8.58 (3.34) 8.05 (3.96) 7.62 (3.67) 7.64 (3.22) 6.59 (4.02) 5.42 (2.23)

SNRV 12.41 (3.03) 12.78 (3.55) 12.38 (2.68) 12.67 (3.37) 13.16 (2.91) 12.57 (3.14) 11.25 (3.05) 11.39 (3.06)



Latency wave V (ms)

Latency wave III (ms) 5

8

r = 0.95 4.5

7

FPP

FPP

r = 0.99

7.5

4

6.5 6 5.5

3.5

5 3.5

4

4.5

5

5

5.5

6

6.5

MAN

Amplitude wave III (µV)

7.5

8

Amplitude wave V (µV)

0.5

0.5

r = 0.91

r = 0.94

0.4

0.4

0.3

0.3

FPP

FPP

7

MAN

0.2 0.1

0.2 0.1

0 0

0.1

0.2

0.3

0.4

0.5

0 0

0.6

0.1

0.2

0.3

MAN

0.4

0.5

0.6

MAN

Fig. 3 – Comparative analysis of the latencies and amplitudes of waves III and V estimated manually by an audiologist and automatically by the FPP method. The linear regression model of the experimental data is compared with the curve FPP = MAN (dotted line).

to present the test ABR signals to the evaluators and ask for the subjective quality. For each level of quality, two ABR signals were presented to the evaluator as reference. The presentation order of the ABR signals was randomized for each test. Fig. 5 shows a screenshot of the computer application for subjective evaluation. This experiment also includes a response validation study carried out by the aforementioned automated methods in

Subject 1 III

Subject 2

V

III

terms of sensitivity and specificity with the same set of ABR signals. The validation of responses by the automated methods was implemented considering a threshold level of quality, which varied in all methods from their lowest estimation of the quality to its greatest value. Automatic evaluations greater or equal to such threshold would be a “positive”, and they would be a “negative” otherwise. These automatic “positive” and “negative” evaluations were compared to an objective

Subject 3

V

III

Subject 4

V

III

Subject 5

V

III

SNR (dB)

SNR (dB)

45 Hz

III: 9.3

III: 8.1

III: 14.3

III: 9.5

III: 6.1

V: 11.0

V: 11.4

V: 14.9

V: 13.4

V: 12.0

55 Hz

III: 8.7

III: 6.6

III: 15.6

III: 2.0

III: 9.0

V: 17.5

V: 8.5

V: 15.7

V: 5.5

V: 15.8

72 Hz

III: 10.7

III: 8.3

III: 14.8

III: 16.1

III: 6.8

V: 17.2

V: 11.6

V: 9.3

V: 12.7

V: 10.6

83 Hz

III: 10.5

III: 6.0

III: 1.8

III: 2.1

III: 4.4

V: 8.7

V: 13.5

V: 6.3

V: 14.4

V: 10.0

100 Hz

III: 10.8

III: 3.9

III: 11.1

III: 5.9

III: 13.9

V: 12.4

V: 13.6

V: 12.7

V: 11.2

V: 12.7

125 Hz

III: 5.2

III: 4.4

III: 5.9

III: 5.1

III: 5.2

V: 10.3

V: 15.7

V: 11.7

V: 14.6

V: 10.9

167 Hz

III: 3.5

III: 7.5

III: 1.7

III: 1.4

III: 1.8

V: 3.7

V: 14.1

V: 8.6

V: 15.0

V: 6.6

250 Hz

III: 5.2

III: 2.9

III: 4.1

III: −4.4

III: 8.0

V: 18.1

V: 14.3

V: 14.4

V: 17.4

V: 7.4

0

2

4

6

Time (ms)

8

10

0

2

4

6

Time (ms)

8

10

0

2

4

6

Time (ms)

8

10

0

2

4

SNR (dB)

V

SNR (dB)

6

Time (ms)

8

10

0

2

4

SNR (dB)

6

8

0.3 μV

10

Time (ms)

Fig. 4 – Examples of ABR signals from five normal hearing subjects obtained at different stimulation rates using the randomized stimulation and averaging (RSA) technique [43]. The parametric peaks adjusted to waves III and V are highlighted on this figure and the automatic quality evaluation provided by the FPP method for each wave is presented.



Table 3 – Mean (and standard deviation in parentheses) of the latencies (L), amplitudes (A), widths (W), and SNRs of waves III and V measured manually on a set of 320 ABR signals obtained from 8 normal hearing subjects at different stimulation rates. Latencies and widths are measured in ms and amplitudes in ␮V. Stimulation rate (Hz) 45 55 72 83 100 125 167 250

LIII 3.73 (0.15) 3.78 (0.11) 3.87 (0.14) 3.91 (0.16) 3.90 (0.14) 4.00 (0.20) 4.16 (0.19) 4.36 (0.37)

LV

LV –LIII

AIII

5.70 (0.23) 5.78 (0.22) 5.90 (0.22) 5.91 (0.40) 6.07 (0.24) 6.21 (0.23) 6.40 (0.29) 6.77 (0.30)

1.97 (0.20) 1.98 (0.18) 2.03 (0.19) 1.97 (0.43) 2.11 (0.20) 2.20 (0.24) 2.22 (0.27) 2.41 (0.21)

0.25 (0.10) 0.23 (0.10) 0.21 (0.08) 0.20 (0.09) 0.17 (0.08) 0.15 (0.07) 0.15 (0.06) 0.12 (0.06)

decision of response validation. This objective decision was made considering the averaged subjective evaluations of the experts greater or equal to 2, which corresponds with the detection of at least the wave V. The sensitivity and specificity parameters for each automated method were estimated at different acceptance thresholds as the true positive rate (TPR: true positives divided by all positives) and as 1-false positive rate (FPR: false positives divided by all negatives) respectively.

3.3.2.

Results

Some examples of ABR signals used for this experiment, including their associated quality evaluation provided by the automatic and subjective methods, are shown in Fig. 6 and Table 4. In this table, FPP is expressed in dB, r is in the range [−1,1], FSP is in absolute value, Cross Corr is in the range [−1,1], and subjective evaluations in the range [0,5]. Signals K and L were obtained without any auditory stimuli, thus no ABR can be detected. Fig. 7A represents the regression analysis between the subjective evaluations provided by five experts and the

AV 0.29 (0.07) 0.28 (0.09) 0.25 (0.07) 0.24 (0.07) 0.22 (0.07) 0.21 (0.05) 0.20 (0.08) 0.17 (0.04)

automatic quality assessment technique based on FPP. The linear regression analysis for each individual subjective evaluation compared to the FPP method is shown in the figure. The correlation coefficient for the regression analysis that considers all subjective evaluations (r = 0.72) is lower in comparison with the mean of the correlation coefficient for the individual evaluations, which suggests that there exists a bias among the evaluations of the experts. On the other hand, the correlation coefficient increases significantly on the regression analysis that considers the average of the subjective evaluations (r = 0.84, Fig. 7B), which remarks that the model is better described with the averaging of a number of individual subjective evaluations. The correlation coefficient for the rest of the automatic methods compared to the averaged subjective evaluations is r = 0.78 for the evaluation based on the correlation coefficient, r = 0.77 for the evaluation based on the FSP expressed in dB, and r = 0.74 for the evaluation based on the cross correlation with a predefined template waveform. The linear regression analysis between the averaged subjective

Fig. 5 – Computer application screenshot used on the subjective evaluation of the quality. Two ABR signals are shown as reference for each quality level. The subjective evaluator is asked to rate the quality for each test ABR between 0 (no ABR) and 5 (excellent quality ABR).



A

G

B

H

C

I

D

J

E

K 0.5 µV

F

L 0

2

4

6

8

10

0

2

Time (ms)

4

6

8

10

Time (ms)

Fig. 6 – Examples of ABR signals of different quality used for test. The signals K and L are obtained without auditory stimulation. The quality evaluation provided for each signal by both automatic and subjective methods is provided in Table 4.

evaluation and the automatic methods based on the correlation coefficient (r), the FSP , and the cross correlation method with a predefined template (Cross Corr) is available as supplementary material (section C). Fig. 8 shows the receiver operating characteristics (ROC) space of a response validation study defined by the false positive rate (FPR), or 1-specificity, and the true positive rate (TPR), or sensitivity, for the automated response validation methods based on fitted parametric peaks (FPP), on the correlation coefficient (r), on the FSP , and on the cross correlation with a predefined template waveform (Cross Corr). This figure shows that the FPP method presents the best results determining the existence of response for all evaluated thresholds, in exception for the thresholds corresponding to FPR evaluations lower than 0.006. The advantage of FPP with the other methods is especially remarkable for low FPR evaluations (lower than 0.1). The FSP method presents better performance than the r and Cross Corr methods for most of the evaluated thresholds. For

FPR evaluations greater than 0.55, the performances of the r, FSP , and Cross Corr methods are very similar.

4.

Discussion

This paper describes in detail and evaluates the fitted parametric peaks (FPP) method, a new approach of automatic quality assessment and peak parameterization based on the use of templates. The use of templates for this purpose was first proposed by C. Elberling in [40]. In his work, a cross correlation method between the ABR signal used for test and a template waveform is described. This method has the limitation of requiring a database of predefined templates for each recording condition, and while a significant match may signify a response, lack of a match do not necessarily means that no response is present, since a response could exist but not match the template [11,40]. Another similar

Table 4 – Evaluation of the quality provided by the automatic evaluation techniques based on FPP, r, FSP , and Cross Corr, by the individual subjective evaluation of the experts (Ev1–Ev5), and by the averaged subjective evaluation (All Ev) for the ABR signals shown in Fig. 6 as examples. ABR A B C D E F G H I J K L

FPP 8.8 10.6 7.6 14.2 7.1 5.8 6.5 4.8 1.4 1.9 1.9 −1.7

r

FSP

0.97 0.99 0.95 0.54 0.70 0.42 0.53 0.61 0.10 0.27 0.40 −0.17

54.1 113.8 12.5 3.6 5.6 2.5 3.7 2.1 1.6 2.1 1.7 0.6

Cross Corr 0.84 0.77 0.86 0.80 0.58 0.61 0.65 0.71 0.64 0.59 0.62 0.36

Ev1

Ev2

Ev3

Ev4

Ev5

All Ev

5 5 5 4 4 3 3 4 0 1 0 0

5 5 4 4 3 4 1 3 2 3 1 0

4 3 3 4 3 3 1 2 1 1 0 0

5 5 4 5 5 3 3 4 1 3 0 0

5 5 5 4 4 3 4 3 2 2 0 0

4.8 4.6 4.2 4.2 3.8 3.2 2.4 3.2 1.2 2.0 0.2 0.0



A) Global and individual subjective evaluations 5

Subjective Evaluation

4

Evaluator 1 (r = 0.80) Evaluator 2 (r = 0.71) Evaluator 3 (r = 0.74) Evaluator 4 (r = 0.75) Evaluator 5 (r = 0.77) All evaluations (r = 0.72)

3

2

1

0

−10

−5

0

5

10

15

10

15

SNR based on Fitted Parametric Peaks (dB) B) Averaged subjective evaluation r = 0.84 5

Subjective Evaluation

4

3

2

1

0

−10

−5

0

5

SNR based on Fitted Parametric Peaks (dB) Fig. 7 – (A) Linear regression analysis for each individual subjective evaluation compared to the automatic evaluation provided by the FPP method. (B) Linear regression analysis for the averaged subjective evaluation. This figure highlights the existing bias among evaluators. The model is better described when an averaged subjective evaluation is considered (r = 0.84).



ROC Space 1 0.9 0.8

TPR (sensitivity)

0.7 0.6 0.5 0.4 0.3 0.2 FPP r Fsp Cross Corr

0.1 0 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

FPR (1 − specificity)

Fig. 8 – ROC space of a response validation study defined by the false positive rate (FPR), or 1-specificity, and the true positive rate (TPR), or sensitivity, for the automated response validation methods based on fitted parametric peaks (FPP), on the correlation coefficient (r), on the FSP , and on the cross correlation with a predefined template waveform (Cross Corr).

template-matching detection algorithm was commercially implemented in the Algo-1 automated evoked response infant hearing screener4 (and successive versions). This detection algorithm is based on the weighting of a number of points in a template waveform according to their relative contribution in identifying a response, and evaluating a test signal in terms of likelihood ratio [53]. Clinical studies carried out by the Algo-1 screener show evidences of a high performance in screening applications [19,53–55]. The approach of the FPP method consists of the search of the latency, width, and amplitude of a parametric peak, similar in morphology to an ABR wave that best fits the most robust waves of the ABR, waves III and V. The parametric peak waveform used as template in the FPP method is commonly known as Mexican hat wavelet, which has been successfully used in different applications of related fields, e.g., [56,57]. The search of the parameters of the fitted peak is computationally optimized to a 1-dimensional search on the width. The optimal latency and amplitude of the parametric peak are directly estimated for a given width. The FPP method described in this paper provides an automatic evaluation of the quality of ABR signals, and parameterizes the most robust waves in terms of amplitude, latency, and width. The performance of the FPP method was evaluated in this study by two experiments. In the first experiment, the latencies and amplitudes of waves III and V were estimated manually by an audiologist and automatically by the FPP

4

Natus Medical Incorporated, San Carlos, CA.

method in ABR signals obtained from eight normal hearing subjects at different stimulation rates. This analysis shows that the FPP method successfully identified all waves III and V. Additionally, the models for latencies and amplitudes of waves III and V as stimulation rate increases are better described when the values are estimated by the FPP method than manually (R2 FPP > R2 MAN in all parameters), which suggests that the FPP method provides more consistent results than the manual procedure, possibly due to the fact that the FPP method bases the estimation of the parameters considering an interval of the response, rather than isolated samples, which makes the FPP method less sensitive to noise. In addition, the results of this experiment show that, despite the difference of a few tens of nanovolts on the estimation of the amplitude of wave V, the FPP method provides an accurate automatic measure of the latencies, amplitudes, and widths of waves III and V, consistent with previous studies. In the second experiment, the performance of FPP was contrasted with the most common automatic quality evaluation procedures: the correlation coefficient (r) [35], the FSP [36], and the cross correlation with a predefined template waveform (Cross Corr) [40]. These automatic quality evaluation methods were compared to a subjective evaluation provided by five experts. The results of this test revealed that although all automatic methods present high correlation coefficients with the averaged subjective assessment, the FPP remains as the method that best approaches an averaged subjective evaluation. Comparing the reliability of the visual judgments provided by the five experts, this test shows, on one hand, that the correlation coefficient is lower when all evaluations are considered in comparison to individual evaluations, and on the other hand, that the correlation coefficient is greater when considering an averaged subjective evaluation. These results suggest that there is an important bias among the evaluators. All individual evaluations present a similar behavior, but a different scale, which evidences that the reproducibility of visual judgments is not high. This conclusion is in accordance with previous studies [33,38,39], and reveals the convenience of using automatic methods. In comparison with the subjective approach, automatic quality assessment methods are uniform, consistent worldwide, and eliminate human inaccuracies. In addition to this, the objective comparison of the aforementioned automated methods in validating ABR signals (Fig. 8) shows that the FPP method presents the best results in most of the thresholds analyzed in the study. The advantages of FPP in research applications are numerous. For instance, the automatic parameterization of the peaks could replace the manual labeling of waves in clinical reports, a tedious task which is usually omitted by the clinical personnel [1]. Furthermore, this functionality could be valuable to provide an automatic ABR interpretation based on response tracking (i.e., analyzing the changes on the morphology of the auditory responses according to a gradual modification of any stimulation setting, such as the intensity level or the stimulation rate). An accurate automatic ABR interpretation might have a significant clinical benefit by helping audiologists on the human decision making [17–23]. The online quality assessment and parameterization of the peaks carried out by FPP could also be appropriate in many real time clinical applications, such as the on-going evaluation of the recorded signal to

Author's personal copy c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 4 ( 2 0 1 4 ) 262–275

automatically stopping averaging, thus eliminating unnecessary recording time [13,14]. In addition to this, the automatic evaluation of the quality of ABR signals could be useful to carry out objective comparisons between the performances of different stimulation methods (RSA [43], QSD [58], CLAD [59,60], etc.) and the effectiveness of different artifact rejection techniques. The FPP method is not defined for clinical applications, such as screening or diagnosing. Screening and diagnosing systems, like the Algo-1 infant hearing screener (Natus Medical Incorporated, San Carlos, CA) [53], are designed to detect waveform abnormalities in very specific recording settings, i.e., nature of the stimuli (clicks, chirps, windowed tones, etc.), polarity, level, rate, hardware equipment, calibration, recording procedure, etc. In screening and diagnosing applications, all parameters involved in the recording process are protocoled and closed, in exception of the subjects. Therefore, screening and diagnosing systems are useful classifying subjects as “normal” (pass) or “pathologic” (fail). The definition of the “pass” criterion requires a strictly protocoled recording procedure (recording system, stimulation and recording settings, etc.) and a clinical study with a large database of explored normal and pathologic subjects. In contrast to these systems, FPP can be used in a wide range of scenarios because it adapts to the normal fluctuations in amplitude, latency, width, and morphology among subjects and recording conditions. These features are appropriate in many research applications. The automatic quality evaluation methods based on the correlation coefficient, the FSP , and FPP present different approaches. First, the correlation coefficient bases the evaluation of the quality on the grade of reproducibility of two consecutive signals. A high positive correlation coefficient would indicate a high quality ABR if both signals are recorded in similar conditions [61]. This method presents the limitation that requires a second ABR signal to perform the test, which doubles the recording time. Additionally, a strong artifact synchronized with the stimulus would lead to an inaccurately high evaluation of the quality. The FSP method bases the evaluation of the quality on the power of the averaged signal and the power of noise across sweeps. This technique requires the evaluation of all recorded sweeps, thus this method cannot be implemented offline unless the EEG is stored (or at least the single point of each sweep). In addition, this technique may present a lack of reliability when evaluating a signal that could not be a response. For instance, this technique would provide a high evaluation index when the ABR is affected by a strong artifact synchronized with the stimulus. Finally, the FPP method approaches the perspective of expert subjective evaluators, rating the grade of identification and quality of the most important waves, does not require the access to the EEG, and provides information regarding the parameterization of the peaks. We believe that since the correlation coefficient method measures the reproducibility of the response, the FSP method measures the level of noise of the recording, and the FPP method evaluates the existence of ABR waves, the use of a combination of all these automatic methods could improve significantly the accuracy in automatic evaluations and provide a better automatic interpretation of ABR signals.

273

Future research could include the search of appropriate template functions that fit the waves of other auditory evoked potentials, such as compound action potentials (CAPs), middle latency responses (MLRs), or late latency responses (LLRs) using the approach of FPP.

5.

Conclusion and significance

A novel automatic method for quality assessment and peak identification of ABR signals, the fitted parametric peaks (FPP), is described and evaluated in this article. The approach of FPP opens a new paradigm in template-matching algorithms, avoiding the need of a database of templates and including additional information regarding the most relevant components of ABR signals. The computational efficiency of the FPP method could be appropriate for its implementation in real time processing applications. The results presented in this article suggest that FPP method presents a high level of accuracy identifying the most important waves of the ABR, and estimating their latency, amplitude, and width. The measure of these parameters with the FPP method seems to be less sensitive to noise than the manual procedure because it considers an interval of the response rather than isolated samples. The automatic identification of the peaks could facilitate the wave labeling process and could be useful to provide an automatic ABR interpretation, with a significant clinical value by helping the operator with the decision making. In comparison with the automatic evaluation techniques based on the correlation coefficient (r), on FSP , and on the cross correlation with a predefined template waveform (Cross Corr), the FPP remains as the method (a) that best approaches a subjective evaluation of the quality, and (b) that provides the best results in the validation of ABR signals in most of the analyzed thresholds. This study has also shown that the subjective evaluations provided by different experts were biased among evaluators, i.e., all evaluators had the same criteria but their scales of assessment were different. This bias can be a problem for the reliability of a subjective evaluation, especially when the evaluator is not an expert. The use of the automatic FPP method described in this paper could be valuable in this context.

Conflict of interest The authors declare no conflicts of interest related to this research work.

Acknowledgments The authors gratefully acknowledge the participation of the voluntary subjects and evaluators involved in this study. This research is granted by the project “Design, implementation and evaluation of an advanced system for recording Auditory Brainstem Response (ABR) based in encoded signalling” (TEC2009-14245), R&D National Plan (2008-2011), Ministry of Economy and Competivity (Government of Spain) and “European Regional Development fund Programme” (2007-2013); by the Granada Excellence Network of Innovation Laboratories - Startup Projects for Young Researchers Programme



(GENIL-PYR 2014), Campus of International Excellence, Ministry of Economy and Competitivity (Government of Spain); and by the grant “University Professor Training Program” (FPU, AP2009-3150), Ministry of Education, Culture, and Sports (Government of Spain).

Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/ j.cmpb.2014.02.015.

references

[1] J.W. Hall, New Handbook of Auditory Evoked Responses, Pearson: Allyn and Bacon, Boston, MA, 2007. [2] C. D’Avanzo, A. Goljahani, G. Pillonetto, G. Nicolao, G. Sparacino, A multi-task learning approach for the extraction of single-trial evoked potentials, Computer Methods and Programs in Biomedicine 110 (2) (2013) 125–136. [3] A.R.D. Thornton, A. Slaven, Auditory brainstem responses recorded at fast stimulation rates using maximum length sequences, British Journal of Audiology 27 (3) (1993) 205–210. [4] S.M. Leung, A. Slaven, A.R.D. Thornton, G.J. Brickley, The use of high stimulus rate auditory brainstem responses in the estimation of hearing threshold, Hearing Research 123 (1–2) (1998) 201–205. [5] H. Zabala-Fernandez, R. Orglmeister, L. Trahms, T.H. Sander, Identification enhancement of auditory evoked potentials in EEG by epoch concatenation and temporal decorrelation, Computer Methods and Programs in Biomedicine 108 (3) (2012) 1097–1105. [6] A. Erenberg, J. Lemons, C. Sia, D. Tunkel, P. Ziring, M. Adams, J. Holstrum, M. McPherson, N. Paneth, B. Strickland, Newborn and infant hearing loss: detection and intervention, Pediatrics 103 (2) (1999) 527–530. [7] Z.D. Jiang, D.M. Brosi, X.M. Shao, A.R. Wilkinson, Maximum length sequence brainstem auditory evoked responses in term neonates who have perinatal hypoxia ischemia, Pediatric Research 48 (5) (2000) 639–645. [8] J.T. Valderrama, A. de la Torre, I. Alvarez, J.C. Segura, A.R.D. Thornton, M. Sainz, J.L. Vargas, A study of adaptation mechanisms based on ABR recorded at high stimulationrate, Clinical Neurophysiology 125 (4) (2014) 805–813. [9] R.F. Burkard, M. Don, The Auditory Brainstem Response, in Auditory Evoked Potentials. Basic Principles and Clinical Application, Lippincott Williams & Wilkins, Baltimore, MD, 2007 (Chapter 11). [10] D.L. Jewett, J.S. Williston, Auditory-evoked far fields averaged from the scalp of humans, Brain 94 (4) (1971) 681–696. [11] C. Elberling, M. Don, Detecting and Assessing Synchronous Neural Activity in the Temporal Domain (SNR, Response Detection), in Auditory Evoked Potentials. Basic Principles and Clinical Application, Lippincott Williams & Wilkins, Baltimore, MD, 2007 (Chapter 5). [12] E. Sejdic, L.A. Lipsitz, Necessity of noise in physiology and medicine, Computer Methods and Programs in Biomedicine 111 (2) (2013) 459–470. [13] M. Don, C. Elberling, Use of quantitative measures of auditory brain-stem response peak amplitude and residual background noise in the decision to stop averaging, Journal of the Acoustical Society of America 99 (1) (1996) 491–499.

[14] O. Ozdamar, R.E. Delgado, Measurement of signal and noise characteristics in ongoing auditory brainstem response averaging, Annals of Biomedical Engineering 24 (6) (1996) 702–715. [15] G. Sparacino, S. Milani, E. Arslan, C.A. Cobelli, Bayesian approach to estimate evoked potentials, Computer Methods and Programs in Biomedicine 68 (3) (2002) 233–248. [16] R.E. Delgado, O. Ozdamar, Automated auditory brainstem response interpretation, IEEE Engineering in Medicine and Biology Magazine 13 (2) (1994) 227–237. ˜ [17] G.G. Gentiletti-Faenze, O. Yanez-Suarez, J.M. Cornejo-Cruz, Evaluation of automatic identification algorithms for auditory brainstem response used in universal hearing loss screening, in: Proceedings of the 25th Annual International Conference of the IEEE Engineering in Medicine and Biology – Proceedings 3, Cancun, Mexico, 2003, pp. 2857–2860. [18] R.E. Eilers, E. Miskiel, O. Ozdamar, R. Urbano, J.E. Widen, Optimization of automated hearing test algorithms: simulations using an infant response model, Ear and Hearing 12 (3) (1991) 191–198. [19] B.S. Hermmann, A.R. Thornton, J.M. Joseph, Automated infant hearing screening using the ABR: development and validation, American Journal of Audiology 4 (2) (1995) 6–14. [20] W.H. McFarland, F.B. Simmons, F.R. Jones, An automated hearing screening technique for newborns, Journal of Speech and Hearing Disorders 45 (4) (1980) 495–503. [21] O. Ozdamar, R.E. Delgado, R.E. Eilers, R.C. Urbano, Automated electrophysiologic hearing testing using a threshold-seeking algorithm, Journal of the American Academy of Audiology 5 (2) (1994) 77–88. [22] K.D. Pool, T. Finitzo, Evaluation of a computer-automated program for clinical assessment of the auditory brain stem response, Ear and Hearing 10 (5) (1989) 304–310. [23] Joint Committee on Infant Hearing, Year 2007 position statement: principles and guidelines for early hearing detection and intervention programs, Pediatrics 120 (4) (2007) 898–921. [24] M. Cebulla, E. Stürzebecher, K. Wernecke, Objective detection of auditory brainstem potentials. Comparison of statistical tests in the time and frequency domains, Scandinavian Audiology 29 (1) (2000) 44–51. [25] E. Stürzebecher, M. Cebulla, Objective detection of auditory evoked potentials. Comparison of several statistical tests in the frequency domain on the basis of near-threshold ABR data, Scandinavian Audiology 26 (1) (1997) 7–14. [26] E. Vannier, O. Adam, P. Karasinski, M. Ohresser, J. Motsch, Computer-assisted ABR interpretation using the automatic construction of the latency–intensity curve, Audiology 40 (4) (2001) 191–201. [27] M.S. John, T.W. Picton, MASTER: a Windows program for recording multiple auditory steady-state responses, Computer Methods and Programs in Biomedicine 61 (2) (2000) 125–150. [28] J. Fridman, E.R. John, M. Bergelson, J.B. Kaiser, H.W. Baird, Application of digital filtering and automatic peak detection to brain stem auditory evoked potential, Electroencephalography and Clinical Neurophysiology 53 (4) (1982) 405–416. [29] F. Chan, F. Lam, P. Poon, W. Qiu, Detection of brainstem auditory evoked potential by adaptive filtering, Medical and Biological Engineering and Computing 33 (1) (1995) 69–75. [30] T. Gronfors, Peak identification of auditory brainstem responses with multifilters and attributed automaton, Computer Methods and Programs in Biomedicine 40 (2) (1993) 83–88. [31] G.C. Galbraith, Enhanced brainstem and cortical evoked response amplitudes: single-trial covariance analysis, Perceptual and Motor Skills 92 (3 (Pt 1)) (2001) 659–672.

Author's personal copy c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 1 4 ( 2 0 1 4 ) 262–275

[32] V. Sundaramoorthy, M.J. Pont, C. Degg, J.A. Cook, A computerized database of ‘normal’ auditory brainstem responses, British Journal of Audiology 34 (3) (2000) 197–201. [33] E. Vannier, O. Adam, J. Motsch, Objective detection of brainstem auditory evoked potentials with a priori information from higher presentation levels, Artificial Intelligence in Medicine 25 (3) (2002) 283–301. [34] J.W. Hall, K.A. Rupp, Auditory brainstem response: recent developments in recording and analysis, Advances in Otorhinolaryngology 53 (1997) 21–45. [35] B.A. Weber, G.L. Fletcher, A computerized scoring procedure for auditory brainstem response audiometry, Ear and Hearing 1 (5) (1980) 233–236. [36] C. Elberling, M. Don, Quality estimation of averaged auditory brainstem responses, Scandinavian Audiology 13 (3) (1984) 187–197. [37] J.T. Valderrama, I. Alvarez, A. de la Torre, J.C. Segura, M. Sainz, J.L. Vargas, A portable, modular, and low cost auditory brainstem response recording system including an algorithm for automatic identification of responses suitable for hearing screening, in: Proceedings of the IEEE/EMBS Special Topic Conference on Point-of-Care HealthCare Technologies (PoCHT), Bangalore, India, 2013, pp. 180–183 (art. no. 6461314). [38] S.A. Arnold, Objective versus visual detection of the auditory brain stem response, Ear and Hearing 6 (3) (1985) 144–150. [39] T. Kakiashvili, W.W. Koczkodaj, M. Woodbury-Smith, Improving the medical scale predictability by the pairwise comparisons method: evidence from a clinical data study, Computer Methods and Programs in Biomedicine 105 (3) (2012) 210–216. [40] C. Elberling, Auditory electrophysiology. The use of templates and cross correlation functions in the analysis of brain stem potentials, Scandinavian Audiology 8 (3) (1979) 187–190. [41] T.W. Picton, D.R. Stapells, K.B. Campbell, Auditory evoked potentials from the human cochlea and brainstem, Journal of Otolaryngology 10 (Suppl. 9) (1981) 1–41. [42] J.L. Stone, M. Calderon-Arnulphi, K.S. Watson, K. Patel, N.S. Mander, N. Suss, J. Fino, J.R. Hughest, Brainstem auditory evoked potentials – a review and modified studies in healthy subjects, Journal of Clinical Neurophysiology 26 (3) (2009) 167–175. [43] J.T. Valderrama, I. Alvarez, A. de la Torre, J.C. Segura, M. Sainz, J.L. Vargas, Recording of auditory brainstem response at high stimulation rates using randomized stimulation and averaging, Journal of the Acoustical Society of America 132 (6) (2012) 3856–3865. [44] J.T. Valderrama, I. Alvarez, A. de la Torre, J.C. Segura, M. Sainz, J.L. Vargas, Educational approach of a BAER recording system based on experiential learning, Technics Technologies Education Management 6 (4) (2011) 876–889. [45] A.R.D. Thornton, M.J. Coleman, The adaptation of cochlear and brainstem auditory evoked potentials in humans, Electroencephalography and Clinical Neurophysiology 39 (4) (1975) 399–406.

275

[46] P.G. Gillespie, U. Muller, Mechanotransduction by hair cells: models, molecules, and mechanisms, Cell 139 (1) (2009) 33–44. [47] M.A. Petoe, A.P. Bradley, W.J. Wilson, Spectral and synchrony differences in auditory brainstem responses evoked by chirps of varying durations, Journal of the Acoustical Society of America 128 (4) (2010) 1896–1907. [48] H. Pratt, H. Sohmer, Intensity and rate functions of cochlear and brainstem evoked responses to click stimuli in man, Archives of Otorhinolaryngology 212 (2) (1976) 85–92. [49] T. Yagi, K. Kaga, The effect of the click repetition rate on the latency of the auditory evoked brain stem response and its clinical use for a neurological diagnosis, Archives of Otorhinolaryngology 222 (2) (1979) 91–96. [50] R.E. Lasky, A developmental study on the effect of stimulus rate on the auditory evoked brain-stem response, Electroencephalography and Clinical Neurophysiology 59 (5) (1984) 411–419. [51] G. Lina-Granade, L. Collet, A. Morgon, B. Salle, Maturation and effect of stimulus rate on brainstem auditory evoked potentials, Brain and Development 15 (4) (1993) 263–269. [52] Z.D. Jiang, Y.Y. Wu, A.R. Wilkinson, Age-related changes in BAER at different click rates from neonates to adults, Acta Paediatrica, International Journal of Paediatrics 98 (8) (2009) 1284–1287. [53] J.G. Peters, An automated infant screener using advanced evoked response technology, Hearing Journal 39 (1986) 25–30. [54] K.J. Doyle, S. Fujikawa, P. Rogers, E. Newman, Comparison of newborn hearing screening by transient otoacoustic emissions and auditory brainstem response using ALGO-2, International Journal of Pediatric Otorhinolaryngology 43 (3) (1998) 207–211. [55] K.J. Doyle, B. Burggraaff, S. Fujikawa, J. Kim, Newborn hearing screening by otoacoustic emissions and automated auditory brainstem response, International Journal of Pediatric Otorhinolaryngology 41 (2) (1997) 111–119. [56] V. Bajaj, R.B. Pachori, Automatic classification of sleep stages based on the time-frequency image of EEG signals, Computer Methods and Programs in Biomedicine 112 (3) (2013) 320–328. [57] J. Widjaja, U. Suripon, Retinal blood vessel detection using wavelet-matched filter, Optical Engineering 52 (3) (2013) 037204. [58] D.L. Jewett, G. Caplovitz, B. Baird, M. Trumpis, M.P. Olson, L.J. Larson-Prior, The use of QSD (q-sequence deconvolution) to recover superposed, transient evoked-responses, Clinical Neurophysiology 115 (12) (2004) 2754–2775. [59] R.E. Delgado, O. Ozdamar, Deconvolution of evoked responses obtained at high stimulus rates, Journal of the Acoustical Society of America 115 (3) (2004) 1242–1251. [60] O. Ozdamar, J. Bohorquez, Signal-to-noise ratio and frequency analysis of continuous loop averaging deconvolution (CLAD) of overlapping evoked potentials, Journal of the Acoustical Society of America 119 (1) (2006) 429–438. [61] S.M. Mason, A.P. Su, R.A. Hayes, Simple online detector of auditory evoked cortical potentials, Medical and Biological Engineering and Computing 15 (6) (1977) 641–647.

Biomed Tech 2014; aop

Joaquin T. Valderrama*, Angel de la Torre, Isaac Alvarez, Jose Carlos Segura, Manuel Sainz and Jose Luis Vargas

A flexible and inexpensive high-performance auditory evoked response recording system appropriate for research purposes Abstract: Recording auditory evoked responses (AER) is done not only in hospitals and clinics worldwide to detect hearing impairments and estimate hearing thresholds, but also in research centers to understand and model the mechanisms involved in the process of hearing. This paper describes a high-performance, flexible, and inexpensive AER recording system. A full description of the hardware and software modules that compose the AER recording system is provided. The performance of this system was evaluated by conducting five experiments with both real and artificially synthesized auditory brainstem response and middle latency response signals at different intensity levels and stimulation rates. The results indicate that the flexibility of the described system is appropriate to record AER signals under several recording conditions. The AER recording system described in this article is a flexible and inexpensive high-performance AER recording system. This recording system also incorporates a platform through which users are allowed to implement advanced signal processing methods. Moreover, its manufacturing cost is significantly lower than that of other commercially available alternatives. These advantages may prove useful in many research applications in audiology. Keywords: auditory brainstem response (ABR); auditory evoked responses (AER); biomedical amplifier; evoked potentials; middle latency response (MLR).

DOI 10.1515/bmt-2014-0034 *Corresponding author: Joaquin T. Valderrama, C/ Periodista Daniel Saucedo Aranda s/n, 18071, Granada, Spain, Phone: +34 958 240 840, Fax: +34 958 240 831, E-mail: [email protected]; [email protected] Joaquin T. Valderrama, Angel de la Torre, Isaac Alvarez and Jose Carlos Segura: Department of Signal Theory, Telematics and Communications, CITIC-UGR, University of Granada, Granada, Spain Manuel Sainz: ENT Service, San Cecilio University Hospital, Granada, Spain; and Department of Surgery and its Specialties, University of Granada, Granada, Spain Jose Luis Vargas: ENT Service, San Cecilio University Hospital, Granada, Spain

Received January 20, 2014; accepted May 5, 2014

Introduction The auditory evoked response (AER) is the electrical activity of the nervous system in response to a stimulus. This electrical activity is characterized by a number of voltage peaks of very low amplitude called evoked potentials, which are generated in different parts of the auditory pathway. These evoked potentials can be classified according to their generator site and the time between stimulus onset and occurrence of the peaks (peak latency), which ranges from 1 ms to 0.5 s. Recording of the AER has been extensively used in human and animal studies for both clinical and research purposes due to its noninvasive nature. The auditory brainstem response (ABR) and the middle latency response (MLR) are AERs generated in the brainstem and in the auditory cortex, respectively [7]. The ABR comprises a number of waves that occur during the first 10 ms from stimulus onset. These ABR waves are identified by sequential Roman numerals as originally proposed by Jewett and Williston [17]. Although up to seven peaks can be seen in the ABR, the most robust waves are I, III, and V. The MLR have latencies from 10 to 60 ms and comprise the components Na, Pa, Nb, and Pb. The longer component of the MLR is usually affected by attention and is difficult to record under sedation. The recording of these signals is commonly used in hospital and clinics worldwide as a hearing screening tool, as well as to detect hearing thresholds and hearing impairments such as vestibular schwannoma and Ménière’s disease. Furthermore, the analysis of the AER may help in understanding the underlying mechanisms involved in the process of hearing [20, 24, 35]. The recording process of these signals requires setting-up a wide range of factors [27]. This paper describes in detail a high-performance, flexible, and inexpensive AER recording system. Although several clinical systems that allow recording of the AER

Brought to you by | Universidad de Granada Authenticated | [email protected] author's copy Download Date | 5/29/14 11:17 AM

2 J.T. Valderrama et al.: A flexible and inexpensive high-performance AER system

already exist, most of them are expensive, they have limited control over most of their parameter settings, and they do not provide access to raw recorded data [1]. In contrast, the AER recording system described in this article gives users full control over the parameter settings. Users can set the intensity level of stimulation, select the number of auditory responses for averaging, use the conventional method of stimulation or any other more advanced techniques, set the stimulation frequency, select the analog-to-digital sampling frequency, choose the order and band-pass cut-off frequencies for digital filters, select the polarity of stimulation and nature of the stimuli (clicks, chirps, tone pips, etc.), or implement advanced artifact rejection techniques. In addition, this system provides access to raw recording data, thus advanced signal processing methods can be implemented offline. The performance of this system was evaluated by conducting five experiments with both real and artificially synthesized ABR and MLR signals recorded at different intensity levels and stimulation rates. The flexibility and inexpensive nature of this high-performance AER recording system may prove it useful in many research applications in audiology.

System architecture System overview Figure 1 outlines the procedure for recording the AER. This process includes the presentation of auditory stimuli and the recording of their corresponding electrical response (sweep) by surface electrodes. A high amplification of this signal is required due to the low amplitude of the

AER (usually < 1 μV). The recorded signal is usually highly contaminated by different types of artifacts, such as myogenic noise related to the muscular activity of the subject, electrical noise derived from the amplifier, electromagnetic and radiofrequency interferences, etc. The conventional method used to reduce the effects of these artifacts and improve the signal-to-noise ratio of the response is the averaging of a large number of sweeps whose corresponding stimuli are periodically presented [4, 8, 38]. This system is battery powered to reduce the artifact generated by the electric power network. The stimulation of the auditory system is conventionally performed by 0.1 ms duration clicks in rarefaction polarity to evoke a synchronous firing of a large number of neurons. However, this system allows the implementation of other stimulus types, such as tone burst, filtered clicks, chirps, noise stimuli, and speech stimuli [13]. The intensity level can be controlled by setting the amplitude of the stimulation signal. A signal composed of a burst of stimuli is generated by the laptop for both stimulation and synchronization purposes. This signal is sent synchronously by the left and right outputs of an analog-to-digital/digital-to-analog (AD/DA) sound card. The right output is connected to the left input for the synchronization of the stimuli. The left output is connected to a pair of insert earphones, through which the stimulation signal excites the auditory system of the subject, thereby generating the AER. This biological signal, plus noise, is recorded by three electrodes placed on the skin at different positions on the head. The electroencephalogram (EEG) recorded by the electrodes is amplified and bandpass filtered. The auditory response after filtering and amplification is recorded synchronously along with the synchronization signal by the right and left inputs of the external AD/DA sound card. The software routines of this system implement the digital signal processing methods

Figure 1 General scheme of the AER recording system.


J.T. Valderrama et al.: A flexible and inexpensive high-performance AER system 3

Figure 2 Picture of the electronics of the amplifier (left) and hardware modules of the AER recording system (right).

necessary to obtain the AER. Figure 2 shows a picture of the electronics of the amplifier (left) and the hardware elements that compose the AER recording system (right). Table 1 presents an estimate of the cost of the elements that make up the AER recording system. This table was built considering the price list of a well-known international electronics supplier. The cost analysis shows that the estimated cost of the elements and materials involved in the AER recording system prototype described in this paper (laptop not included) is about 950 USD.

Hardware specifications Amplifier The electronic schematic of the amplifier is shown in Figure 3. The amplifier is composed of four stages: Table 1 Estimated cost of the elements that compose the AER recording system. Element

a

Amplifier electronics Electrodes and electrolytic paste Etymotic ER·3A insert earphones External AD/DA sound card TOTAL a

Rough cost 200 USD 200 USD 500 USD 50 USD 950 USD

Amplifier electronics include semiconductor elements, integrated circuits, connectors, PCB card, box, batteries, and battery holders.

preamplification, band-pass filtering, amplification, and active ground circuitry. Preamplification provides a moderate gain to avoid saturation in later stages. This stage is done using the instrumental amplifier INA128 (Texas Instruments Inc., Dallas, TX, USA). This differential amplifier was chosen because of its high common mode rejection ratio (CMRR), low power, low noise ( 8 nV / Hz ), and easy control of the gain. Band-pass filtering removes the frequencies out of the scope of the AER, amplifying only the band of interest. This stage comprises four secondorder Sallen-Key filters (2 × high pass and 2 × low pass). The values of the resistors and capacitances that implement the filtering stage define the bandwidth of the amplifier. The bandwidth of the amplifier must be selected considering the characteristic frequencies of each AER. Table 2 shows the characteristic bandwidth for recording ABR and MLR signals, along with suggested values of resistors and capacitances that implement the high pass and low pass filtering stages of the amplifier. These analog filters insert a phase distortion on the recorded signal that must be adjusted by software. This phase shift is 560 μs for the ABR amplifier and 80 μs for the MLR amplifier. The amplification stage after filtering sets the required level of amplitude on the EEG to be recorded by the analog-to-digital converter. The active ground circuit is designed to reduce the common mode voltage of the recorded signal. The electric field generated by the electric network can induce a common mode voltage on the subject. This common mode voltage is amplified, inverted, and inserted back to the subject by the



Figure 3 Electronic circuit diagram of the amplifier.

Table 2 Frequency bandwidth of different AERs and suggested values of resistors and capacitances that implement the high-pass and lowpass filtering stages of the amplifier. Evoked response

Bandwidth

ABR MLR

[150–3500] Hz [0.5–3500] Hz

High pass filter

R1-H

R2-H

C1-H

33 kΩ 1 MΩ

33 kΩ 1 MΩ

47 nF 470 nF

active ground circuit, thereby significantly reducing the common mode voltage on the subject. The operational amplifiers OPA227 (Texas Instruments Inc., Dallas, TX, USA) used in this circuit were chosen because of their very low noise voltage ( 3 nV / Hz ), high CMRR (130 dB), and high precision. The Bode diagrams on Figure 4 show the bandwidth and the phase shift of the amplifiers for ABR and MLR signals. The gain of the amplifier reaches the value GA = 20,000 (86 dB) for the band-pass frequencies,

2

10

3

10

4

100 90 80 70 60 50 40 30 20

10

2

10

R1-L

22 nF 470 nF

6.8 kΩ 6.8 kΩ

0

−π

3

10

4

10

0

−π 2

10

3

10 Frequency (Hz)

4

10

2

10

R2-L

R1-H

R1-H

6.8 kΩ 6.8 kΩ

4.7 nF 4.7 nF

10 nF 10 nF

with a filter slope of 24 dB/oct. Figure 5 presents a linearity analysis for the ABR amplifier. This figure represents a 10 ms sinusoidal signal inserted on the amplifier (input signal) versus its corresponding output signal. The slope of this curve represents the gain of the amplifier (86 dB). The amplitude of the input signal was chosen to obtain a slightly saturated output signal. The frequency of the input signal was set to 1087 Hz to obtain an output signal with phase distortion zero. This analysis suggests that

π Phase (rad)

π Phase (rad)

C2-H

MLR recording system

Gain (dB)

Gain (dB)

ABR recording system

100 90 80 70 60 50 40 30 20

Low pass filter

3

10 Frequency (Hz)

4

10

Figure 4 Bode diagram of the amplifier.



10 8 6 Output signal (V)

4 2 0 −2 −4 −6 −8 −10 −600

−400

−200

0

200

400

600

Input signal (μV)

Figure 5 Linearity analysis of the input signal versus output signal of the ABR amplifier.

the behavior of the amplifier is especially lineal when the input signal is in the range [-0.3 +0.3] mV, a common situation considering that the recorded EEG does not usually exceed 50 μV [13]. Thus, the dynamic range of the amplifier is 600 μv. The consumption of this circuit is 28.2 mA, which allows the device an operating time of more than 6 h for standard rechargeable 9 V batteries. The safety of the subject under exploration is assured, on one hand, by the battery-powered nature of the system, which prevents any possible electrical shock derived from the electrical network; and, on the other hand, by the 1 MΩ resistor that connects the active ground electrode to the subject, which limits the leakage current introduced to the subject to 9 μA, meeting the electrical safety requirements of the international standard IEC 60601-1 [21].

Electrodes Electrodes transform ionic currents (the mechanism of conduction of bioelectrical signals on tissues) into electrical currents that conduct the evoked potentials from the subject to the recording system. Given that the electrodes are the first components of signal recording, the noise level generated at them should be minimized. The electrodes typically used in AER recording to reduce contact potential are silver coated with silver chloride (Ag/AgCl) surface electrodes, which consist of a silver conductor (electrode) immersed into a silver chloride salt dissolution (electrolyte). Electrolytic paste is used as a means of union between the electrode and the skin to reduce contact electrode impedance. The contact impedance of the junction

between the scalp and the electrodes should be kept as low as possible to minimize the magnitude of induced electromagnetic artifacts and to reduce the capacitive coupling effects of the electrode cables and external power lines [13, 27]. This contact impedance can be reduced by softly scraping the skin with alcohol or other cleansing agents. The electrode-skin contact impedance can be measured either by using any commercially available alternating-current impedance meter or by implementing the circuit diagram of any impedance meter described in the literature, e.g., [11, 12]. Impedances lower than 5 kΩ at the working frequencies can be considered acceptable. The electrode impedance should be balanced to avoid common mode artifacts. The placement of the electrodes can be done in accordance with the standard positions defined by the International 10-20 and 10-10 Systems [16, 19]. Active, ground, and reference electrodes can be placed on the high forehead (Fz), low forehead (Fpz), and ipsilateral mastoid (TP9/TP10), respectively, as shown in Figure 1. Active and reference electrodes are connected to the differential inputs of the amplifier. The ground electrode connects the active ground input of the amplifier.

Analog-to-digital conversion Analog-to-digital conversion is performed by an external sound card connected to the laptop through the USB port. This device presents the advantages of simplicity and better performance compared to most sound cards integrated on laptops. Table 3 shows a summary of the features of the AD/DA sound card. The number of bits of quantization and the sampling rate can be controlled by the user. The amplitude precision of the analog-to-digital conversion is determined by the number of bits of quantization. With regard to the recording of ABR signals, the analog-to-digital converter should be able to measure within the range of 2 nV (10% precision of a standard 20 nV amplitude of a wave II) to 200 μV (the highest expected recorded level of an EEG) – a ratio of 100,000, corresponding to a dynamic range of 100 dB. Considering that an AD/ Table 3 Features of the AD/DA sound card. Feature

Sampling rate Input range Output range Bits of quantization Quantization step


Value 25 kHz -3 V/+3 V -2.5 V/+2.5 V 16 91.55 μV


DA of n bits has a dynamic range of 6·n dB, the number of bits of the AD/DA required for it to be able to record ABR with a precision of 10% is about 16 bits. In addition, sweeps averaging increases the precision of the measure, reducing the quantization noise [27]. Therefore, the use 16 bits of quantization is enough to record AER with sufficient precision. The sampling rate must be greater than twice the highest frequency component present in the signal to prevent aliasing [22]. However, the low-pass filters of the filtering stage in the amplifier just attenuate (not eliminate) the frequency components greater than the cutoff frequency. The aliasing errors from all frequency components could be prevented only when the sampling rate is set to twice the frequency at which the filter attenuates the signal by more than the dynamic range of the AD/DA. Considering a standard AD/DA converter, the frequency at which this attenuation occurs is f ′ = fc ⋅ 2 ( D− 3 )/ S , where fc is the cutoff frequency, D is the dynamic range of the AD/ DA in dB, and S is the slope in dB per octave [27]. Therefore, to avoid even 1-bit aliasing errors, the sampling rate (fs) must be fs = 2⋅ fc ⋅ 2 ( D− 3 )/ S . Given that the AER recording system described in this article includes an anti-aliasing filter with a cutoff frequency of 3000 Hz and a steep slope of 24 dB per octave used in conjunction with a 16-bit AD/ DA, the sampling rate must be over 22,982 samples per second to avoid all aliasing errors. Hence, a sampling rate of 25 kHz could be appropriate to avoid all aliasing errors and at the same time prevent the undesired effects of oversampling.

Transducer Earphones provide stimulation to the auditory system of the subjects by transducing the electrical energy of the stimulation signal into acoustical energy (sound). The tubal insert earphones Etymotic ER3A (Etymotic Research, Inc., Elk Grove Village, IL, USA) were chosen for this application because of their flat response to a wide band of frequencies, their isolation from external noise, and their fast response to typical click stimuli, which enables a synchronous firing of inner hair cells [13]. Other standard earphones, such as the Telephonics TDH-39, -49, -50 (Cadwell Laboratories, Inc., Kennewick, WA, USA), can also be used.

Software specifications The software modules involved in the AER recording process are presented in Figure 6. The first step in data

Figure 6 Diagram of the software modules.

acquisition is generation of the stimulation signal. The conventional stimulation technique consists of the presentation of stimuli with a constant interstimulus interval (ISI) greater than the averaging window to avoid overlapping responses [4]. Other more advanced methods, such as maximum length sequences (MLS) [9], continuous loop averaging deconvolution (CLAD) [5, 23], quasiperiodic sequence deconvolution (QSD) [18], least-squares deconvolution (LS) [2, 3], and randomized stimulation and averaging (RSA) [28], can also be implemented to obtain AER at high stimulation rates. The stimulation of the auditory system is typically performed by 0.1 ms duration clicks in rarefaction polarity to evoke a synchronous firing of a large number of neurons [13]. Other types of stimuli, such as tone bursts, filtered clicks, paired clicks, plops, chirps, modulated tones, stimulus trains, noise stimuli, and speech stimuli, can also be implemented. The parameters type of stimuli, intensity level, clicks duration, clicks polarity, stimulation rate, and number of recorded sweeps can be controlled in this module. The “Stimulation & Recording” module consists of (a) the synchronous reproduction of the stimulation signal and (b) the synchronous recording of the stimulation signal and the digitized EEG. In this step, the user has control over the number of quantization bits and the sampling rate. The “Scaling” module functions to convert the recorded signal (AX) into its corresponding value in microvolts at the electrodes. Considering that GA is the gain of the amplifier for the band-pass frequencies and GS is the gain of the AD/DA, the scaled value in microvolts at the elec1 1 trodes is Ascaled ( µV ) = AX × × × 10 6 . The values of GA GS GA and GS are estimated during the calibration process, which



is described in the next section. The “AER enhancement” module incorporates algorithms, such as digital filtering and artifact rejection techniques, to increase the quality of the response. The “Synchronization” module uses the recording of the stimulation signal as trigger to determine the samples at which stimuli occur. The “AER calculation” module runs the necessary algorithms to obtain the AER according to the method used in the stimulation process. This software module also compensates for the phase distortion inserted by the analog filters of the amplifier on the recorded AER. Finally, the “Storage” module saves the raw data, the processed variables, and other important parameters into a file on the database. The parameters involved in the process of recording AERs can be managed from a graphical user interface. The structure of this multimedia platform can be designed according to the specific requirements of the users. Figure 7 shows an example of an interactive front-end of the AER recording system, in which the user has the control over recording parameters, such as the ISI of the stimulation sequence, the number of recorded sweeps, the intensity level, and the duration of the click. This platform also allows the use of specific signal processing techniques, such as digital filtering, frame rejection, and digital blanking, to obtain higher quality signals. Additional information, such as the number of accepted and rejected frames, the

acceptance ratio, and the recorded EEG, is also provided. In this example, the AER, as well as a history of previously recorded signals, is shown in a graph. An example of software routine that implements the recording of AER using the conventional method is available in MATLAB (The Mathworks, Inc., Natick, MA, USA) code as supplementary material (Supplementary data).

Calibration Calibration of GA and GS The calibration process consists of estimating the values of the gain of the amplifier for the band-pass frequencies (GA) and the gain of the AD/DA (GS) to perform a correct scaling of the recording signal. The value of GA can be estimated directly from the Bode diagram of the amplifier. The value of GS is related to the intensity level of the input line of the AD/DA sound card. This parameter can be configured from the audio settings of the laptop. Medium intensity level is recommended to avoid possible nonlinearities. The value of GS can be estimated by correlating a recorded signal whose maximum amplitude in volts is known (Vhi) with its corresponding value of the recorded signal (Xhi), GS = Vhi/Xhi.

Figure 7 Interactive front-end of the AER recording system. This multimedia platform gives the user full control over all parameters involved in the AER recording process.



Calibration of the intensity level The calibration of the intensity level consists of measuring the stimulus magnitude necessary for providing an accurate and uniform evaluation of the evoked responses. Standard audiometric calibration methods include dB normal hearing level (nHL) and dB sound pressure level (SPL) [6]. The intensity level 0 dB nHL represents the hearing threshold for normal hearing subjects. This intensity level can be established as the mean value of the intensity level at which stimuli are just detectable in a set of 15–20 subjects with no auditory dysfunction (normal hearing subjects) [13]. The intensity level of a stimulus in ⎛ P ⎞ terms of dB SPL is estimated as 20 × log10 ⎜ x ⎟ , being Px ⎝ Pref ⎠ the pressure of the stimulus and Pref the reference pressure, whose typical reference value is 20 μPa. A complete description of the procedure to calibrate the reference zero is described by the international standard ISO 389 [15, 26]. In this system, the calibration of the stimuli is performed according to the aforementioned international standard. The intensity level can be controlled by the user through the output voltage of the stimulation signal. Given Vref as the amplitude voltage of a stimulation signal that presents an intensity level of 0 dB nHL, the amplitude voltage necessary to present an intensity level of X dB nHL can be obtained using the formula VX =Vref × 10 X / 20 .

Scalability The use of multiple-channel systems may be required in certain research applications, e.g., the use of binaural stimulation for simultaneous screening in both ears, the use of contralateral masking to ensure monaural stimulation, and the simultaneous screening of ABR and electrocochleography (ECochG) [25]. The AER recording system described in this system is scalable. A multichannel version of this system can be set up using an AD/DA converter of multiple channels and multiple units of the amplifier. Considering that the price of a standard 4-channel AD/DA sound card is about 150 USD, and that the estimated manufacturing cost of an amplifier unit is about 200 USD, the total cost of implementing a 4-channel AER recording system would be about 1250 USD.

Assessment The performance of the AER recording system described in this article was evaluated by conducting five experiments

on one normal hearing subject (#S1: male, 28 year). The subject explored in these experiments was informed about the experimental procedure and possible side effects of the test, and the subject gave consent for the use of the data. The calibration of the intensity level was performed according to the international standard ISO 389 [15, 26]. The equivalent 0 dB nHL corresponds to 36.4 dB SPL. The recording procedure of these experiments was approved by the Clinical Research Ethics Committee of the San Cecilio University Hospital and by the Human Research Ethics Committee of the University of Granada (Reference No. 826014263-14263-4-9), in accordance with the Code of Ethics of the World Medical Association (Declaration of Helsinki) for experiments involving humans. Additionally, this section introduces an outline of related research activities performed using the AER recording system described in this paper. Experiment 1 was designed to simulate the recording of ABR and MLR signals and assess the performance of the AER recording system. The ABR and MLR signals used in this experiment (original pseudopotentials) were obtained from #S1 using 10,000 click stimuli presented at a rate of 33 Hz for ABR and 3.3 Hz for MLR at an intensity level of 70 dB nHL. A burst of 10,000 pseudopotentials was digitally synthesized for each type of signal. The amplitude of both signals was reduced by a voltage divider to obtain signals of 0.2 μV for ABR and 0.5 μV for MLR. The burst of low-amplitude pseudopotentials was amplified, recorded by the AD/DA sound card, and digitally processed according to the recording procedure described in System architecture. Figure 8 shows the original and recorded pseudopotentials for both the ABR and the MLR signals. The most important components of these signals are marked on the figure. This figure shows that the AER recording system described in this article can be used to obtain signals similar in morphology and amplitude to ABR and MLR given that the major components of these signals can be easily identified, they remain on the same latency, and they present similar amplitude. Experiment 2 was devised to analyze the effects of noise reduction through sweeps averaging. Figure 9 shows the ABR and the MLR signals obtained from #S1 at different numbers of averaged sweeps. The stimuli used on this experiment were clicks presented at 70 dB nHL at a stimulation rate of 33 Hz for ABR and 8 Hz for MLR. This figure shows that the quality of the AER increases with the number of averaged sweeps. The main waves of these signals start to be identified with at least 500 sweeps. The recordings obtained with 20,000 sweeps, especially for MLR signals, were of higher quality but they required a longer test time. About 2000 sweeps may be appropriate to



Figure 8 Recording of low-amplitude digitally synthesized signals similar in morphology to ABR and MLR potentials.

robust component, and in this experiment it can be clearly identified up to 15 dB nHL. These results are in accordance with those reported in the literature [14, 17]. Experiments 4 and 5 were set up to analyze the effects of stimulation rate on the morphology of the ABR and the MLR signals, respectively. Figure 11 shows the ABR signals from #S1 obtained at stimulation rates of up to 250 Hz using the RSA [28], QSD [18], and conventional (CONV) techniques [4]. All recordings were obtained using 5000 averaged sweeps stimulated with clicks at 70 dB nHL. The amount of jitter used in the stimulation sequences for both the RSA and the QSD was 4 ms. The jitter of a stimulation sequence measures the grade of dispersion of the ISI compared to a periodical presentation of the stimuli, i.e., the ISI of stimuli presented at a rate of 25 Hz with a jitter of 4 ms would vary between 38 and 42 ms.

reach a compromise between recording time and quality of the recordings. However, AER recordings obtained with a larger number of averaged sweeps can be interesting in certain applications, such as in the study of neural adaptation, that require the analysis of high-quality AER and do not impose significant restrictions on the recording test time [35]. Experiment 3 was developed to evaluate the influence of intensity level on the morphology of ABR signals. Figure 10 shows the ABR signals from #S1 obtained at intensity levels of stimulation that vary from 5 to 80 dB nHL, in steps of 5 dB. For each ABR signal, 5000 sweeps were recorded. Waves I, III, and V are labeled on the ABR signal obtained at 80 dB nHL. This experiment shows that the amplitude of the most relevant waves decreases and their corresponding latency increases as the stimulation intensity level decreases. Wave V remains as the most

MLR

ABR III

Number of averaged sweeps

I

V

V

20,000

20,000

5000

5000

2000

2000

1000

1000

500

500

200

200

100

Pa

Na

Pb

Nb

100 1μV

0.3μV 50

50

0

2

4 6 Time (ms)

8

10

0

20

40 60 Time (ms)

80

100

Figure 9 Influence of the number of averaged sweeps on the quality of the ABR and MLR signals.



Both the RSA and QSD techniques are valid methods to obtain ABR signals at very high stimulation rates ( > 100 Hz). Waves I, III, and V can be clearly identified in all recordings, although the ABR signal obtained with QSD at 250 Hz is slightly noisier. This figure shows the normal changes on the morphology of the ABR as stimulation rate increases: the amplitude of the waves decreases and the latencies increase, with a deeper shift on the most central waves. Figure 12 shows the MLR signals from #S1 obtained at stimulation rates from 8 to 125 Hz obtained with the RSA technique with a jitter of 16 ms, using click stimuli presented at 70 dB nHL. The V, Na, Pa, Nb, and Pb components can be identified at all stimulation rates. These components are labeled on the MLR signal obtained at 125 Hz. The MLR signal obtained at 40 Hz presents a resonance, in which the Na, Pa, Nb, and Pb components are in phase (occurring at the same time relative to the stimulus) and become superimposed. This phenomenon is generally known as 40-Hz eventrelated potential (ERP) and was first described by Galambos et al. in 1981 [10]. The 40-Hz ERP presents advantages for the estimation of the auditory threshold due to its large amplitude (usually > 1 μV). In addition to these five experiments, the AER recording system described in this paper has been successfully used in related research activities. This system was used to develop (a) the RSA method, a technique that allows the recording of

ABR V

III I 80 75 70 65 60 Intensity level (dB nHL)

55 50 45 40 35 30 25 20 15 10

0.3 μV

5 0

2

4

6

8

10

Time (ms)

Figure 10 ABR signals obtained at different intensity levels of stimulation.

RSA III


I

CONV

QSD

V

III

I

V

III

I

45

45

45

56

56

56

71

71

71

83

83

83

100

100

100

125

125

167

167

250

250

V

0.3 μV

0

2

4 6 8 Time (ms)

10

0

2

4 6 8 Time (ms)

10

0

2

4 6 8 Time (ms)

10

Figure 11 ABR signals recorded at different stimulation rates using the randomized stimulation and averaging (RSA), the quasiperiodic sequence deconvolution (QSD), and the conventional (CONV) methods.



MLR V Pa

Pb

125 Nb

Na


100

67

40

20 0.6 μV 8

0

20

40

60

80

100

Time (ms)

Figure 12 MLR signals obtained at different stimulation rates using the randomized stimulation and averaging (RSA) technique.

AER at high rates [28]; (b) the separated response method, which allowed for the first time the study of the fast and slow mechanisms of adaptation in humans [30, 35]; (c) the fitted parametric peaks method, which provides an automatic evaluation of the quality of ABR signals and a parameterization of the most important waves in terms of amplitude, latency, and width [31, 34]; (e) studies to test whether or not high stimulation rates could save recording time [29, 36]; (f) an automatic auditory response detection paradigm based on response tracking [37]; (g) a study of the effects of averaging and deconvolution in ABR and MLR signals using the RSA method [32]; and (h) a deconvolution method based on randomized stimulation using artifact rejection methods in the frequency domain [33].

Discussion This paper provides a full description of a flexible and inexpensive high-performance AER recording system. The system described in this article includes an amplifier, an external sound card that acts as an AD/DA converter of

two I/O channels, electrodes, cables and connectors, and a laptop with software modules. The software modules run the algorithms for the stimulation sequence generation, the production of the stimuli and the recording of the sweeps, the scaling of the recorded EEG, the synchronization of the sweeps with their associated stimuli, the processing of data according to the specific stimulation method to obtain the AER (CONV, MLS, QSD, CLAD, LS, RSA, etc.), and finally, the storage of the EEG and the AER into a file. The open nature of this system provides the flexibility required in many research applications. Almost every parameter involved in the AER recording process can be defined and controlled. For instance, this system allows the user full control over parameters such as the nature, duration and polarity of stimuli; the number of averaged sweeps; the intensity level; and the stimulation rate. The software platform of this system allows the implementation of advanced stimulation methods, such as RSA and QSD, that permit the recording of AER signals at high rates of stimulation, digital filtering to enhance the quality of the recordings, and the use of artifact rejection methods. In addition, the recording of the raw EEG may be of interest to implement advanced signal processing methods offline. Furthermore, the scalability of the system allows the implementation of a multiple-channel design, which may be useful in certain research applications, such as the use of binaural stimulation for simultaneous screening in both ears, the application of contralateral masking to ensure monaural stimulation, and the simultaneous screening of ABR and ECochG [25]. The performance of this system was evaluated by conducting five experiments that included (a) the recording of artificially synthesized ABR and MLR signals (pseudopotentials), (b) the recording of real ABR and MLR signals of different quality using a varying number of averaged sweeps, (c) the analysis of the influence of intensity level on the morphology of the ABR signals, (d) and the study of the effects of stimulation rate on the morphology of the ABR and MLR signals. Some of the results obtained in these experiments, such as the ABR signal obtained at 250 Hz and the MLR signal recorded at 125 Hz (experiments 4 and 5), are especially remarkable. In addition to these experiments, the AER recording system proposed in this article has been proved to be effective in several previous studies, e.g., this architecture was used (a) to develop the RSA method and compare its performance with the QSD technique through ABR signals recorded from eight subjects at different stimulation rates [28]; (b) to study the fast and slow mechanisms of adaptation in humans by analyzing the morphology of ABR signals obtained with the separated responses methodology [30, 35]; (c) to develop and evaluate different



approaches to automatic quality assessment and response detection methods [31, 34, 37]; (d) to conduct a study to test whether or not high stimulation rates could save recording time [29, 36]; (e) to analyse the effects of adaptation and deconvolution of ABR and MLR signals with RSA [32]; and (f) to develop a method that allows the deconvolution of overlapping responses with randomized stimulation using frequency domain-based artifact rejection methods [33]. The results of the experiments, along with the results obtained in the aforementioned studies [28–37], indicate that the AER recording system described in this article can be efficiently used to record ABR and MLR signals under different recording conditions. Although several clinical devices for recording AERs already exist, most of them are expensive and they suffer from a lack of flexibility because they are designed for specific applications (e.g., hearing threshold estimation). Commercial systems designed for research applications are more flexible than the aforementioned clinical devices. However, the flexibility of these systems is limited by the performance of their associated software, and their acquisition price is usually high because it includes not only the cost of the materials, but also costs derived from marketing, distribution, technical support, profit margin, etc. In contrast, the estimated cost for the implementation of a prototype of the AER recording system, including circuitry, connectors, box, external AD/DA sound card, the Etymotic ER·3A insert earphones, electrodes, and cables (laptop not included), is < 1000 USD. The inexpensive and flexible nature of the high performance AER recording system described in this article may prove useful in several research applications in audiology.

Conclusion This article describes in detail the hardware and software elements of a high-performance AER recording system. The performance of this system was assessed by conducting five experiments with both real and artificially synthesized ABR and MLR signals under different recording conditions. The flexibility and inexpensive nature of this high-performance AER recording system may prove useful in several research applications in Audiology.

Supplementary file Supplementary file 1: Example of the MATLAB routine that implements the recording of AER using the conventional method.

Conflict of interest statement There are no conflicts of interest associated with this research article. The authors have no financial involvement or interest with any organization or company about subjects or materials discussed in the paper. Acknowledgments: This research has been supported through a grant by the project “Design, implementation and evaluation of an advanced system for recording Auditory Brainstem Response (ABR) based in encoded signaling” (TEC2009-14245), R&D National Plan (2008-2011), Ministry of Economy and Competivity (Government of Spain) and “European Regional Development fund Programme” (2007-2013); by the “Granada Excellence Network of Innovation Laboratories-Startup Projects for Young Researchers Programme” (GENIL-PYR 2014), Campus of International Excellence, Ministry of Economy and Competitivity (Government of Spain); and by the grant “Formación de Profesorado Universitario” (FPU) (AP20093150), Ministry of Education, Culture, and Sports (Government of Spain).

References [1] Bahmer A, Peter O, Baumann U. Recording of electrically evoked auditory brainstem responses (E-ABR) with an integrated stimulus generator in Matlab. J Neurosci Meth 2008; 173: 306–314. [2] Bardy F, Dillon H, Van Dun B. Least-squares deconvolution of evoked potentials and sequence optimization for multiple stimuli under low-jitter conditions. Clin Neurophysiol 2014; 125: 727–737. [3] Bardy F, Van Dun B, Dillon H, McMahon CM. Deconvolution of overlapping cortical auditory evoked potentials recorded using short stimulus onset-asynchrony ranges. Clin Neurophysiol 2014; 125: 814–826. [4] Dawson GD. A summation technique for the detection of small evoked potentials. Electroencephalogr Clin Neurophysiol 1954; 6: 65–84. [5] Delgado RE, Özdamar O. Deconvolution of evoked responses obtained at high stimulus rates. J Acoust Soc Am 2004; 115: 1242–1251. [6] Durrant JD, Boston JR. Stimuli for auditory evoked potential assessment. In: Sabatini P, Klinger AM, Ajello JP, editors. Auditory evoked potentials. Basic principles and clinical application. Baltimore, MD: Lippincott Williams & Wilkins 2007: 42–72. [7] Eggermont JJ. Electric and magnetic fields of synchronous neural activity. In: Sabatini P, Klinger AM, Ajello JP, editors. Auditory evoked potentials. Basic principles and clinical application. Baltimore, MD: Lippincott Williams & Wilkins 2007: 2–21. [8] Elberling C, Don M. Detecting and assessing synchronous neural activity in the temporal domain (SNR, response detection). In: Sabatini P, Klinger AM, Ajello JP, editors. Auditory evoked potentials. Basic principles and clinical application. Baltimore, MD: Lippincott Williams & Wilkins 2007: 102–123.



[9] Eysholdt U, Schreiner C. Maximum length sequences: A fast method for measuring brain-stem-evoked responses. Audiol 1982; 21: 242–250. [10] Galambos R, Makeig S, Talmachoff PJ. A 40-Hz auditory potential recorded from the human scalp. Proc Natl Acad Sci USA 1981; 78: 2643–2647. [11] Gravaill N. A simple battery-operated a.c. impedance meter. Med Biol Eng Comput 1978; 16: 339–340. [12] Grimnes S. Impedance measurement of individual skin surface electrodes. Med Biol Eng Comput 1983; 21: 750–755. [13] Hall JW. New handbook of auditory evoked responses. 1st ed. Boston, MA: Pearson Allyn and Bacon 2007. [14] Hecox K, Galambos R. Brain stem auditory evoked responses in human infants and adults. Arch Otolaryngol 1974; 99: 30–33. [15] ISO 389-x. Acoustics – Reference zero for the calibration of audiometric equipment – Part 1-9. International Organization for Standard. [16] Jasper HH. The ten-twenty electrode system of the International Federation. Electroencephalogr Clin Neurophysiol 1958; 10: 371–375. [17] Jewett DL, Williston JS. Auditory-evoked far fields averaged from the scalp of humans. Brain 1971; 94: 681–696. [18] Jewett DL, Caplovitz G, Baird B, Trumpis M, Olson MP, LarsonPrior LJ. The use of QSD (q-sequence deconvolution) to recover superposed, transient evoked-responses. Clin Neurophysiol 2004, 115: 2754–2775. [19] Jurcak V, Tsuzuki D, Dan I. 10/20, 10/10, and 10/5 systems revisited. Their validity as relative head-surface-based positioning systems. Neuroimage 2007; 34: 1600–1611. [20] Lasky RE. Rate and adaptation effects on the auditory evoked brainstem response in human newborns and adults. Hearing Res 1997; 111: 165–176. [21] (2012) Medical Electrical Equipment – Part 1: General requirements for basic safety and essential performance. International Electrotechnical Commission IEC 60601-1. [22] Nyquist H. Certain factors affecting telegraph speed. Bell Syst Tech J 1924; 3: 324–346. [23] Ozdamar O, Bohorquez J. Signal-to-noise ratio and frequency analysis of continuous loop averaging deconvolution (CLAD) of overlapping evoked potentials. J Acoust Soc Am 2006; 119: 429–438. [24] Ozdamar O, Bohorquez J, Ray SS. Pb(P1) resonance at 40 Hz: effects of high stimulus rate on auditory middle latency responses (MLRs) explored using deconvolution. Clin Neurophysiol 2007; 118: 1261–1273. [25] Reid A, Thornton ARD. The effects of contralateral masking upon brainstem electric responses. Brit J Audiol 1983; 17: 155–162. [26] Richter U, Fedtke T. Reference zero for the calibration of audiometric equipment using ‘clicks’ as test signals. Int J Audiol 2005; 44: 478–487. [27] Thornton ARD. Instrumentation and recording parameters. In: Sabatini P, Klinger AM, Ajello JP, editors. Auditory evoked potentials. Basic principles and clinical application. Baltimore, MD: Lippincott Williams & Wilkins 2007: 73–101.

[28] Valderrama JT, Alvarez I, de la Torre A, Segura JC, Sainz M, Vargas JL. Recording of auditory brainstem response at high stimulation rates using randomized stimulation and averaging. J Acoust Soc Am 2012; 132: 3856–3865. [29] Valderrama JT, Alvarez I, de la Torre A, Segura JC, Sainz M, Vargas JL. Reducing recording time of brainstem auditory evoked responses by the use of randomized stimulation. Newborn Hearing Screening Congress, June 5–7 2012, Cernobbio (Como Lake), Italy. [30] Valderrama JT, Alvarez I, de la Torre A, Segura JC, Sainz M, Vargas JL. A preliminary study of the short-term and long-term neural adaptation of the auditory brainstem response by the use of randomized stimulation. Adults Hearing Screening Congress, June 7–9 2012, Cernobbio (Como Lake), Italy. [31] Valderrama JT, Alvarez I, de la Torre A, Segura JC, Sainz M, Vargas JL. A portable, modular, and low cost auditory brainstem response recording system including an algorithm for automatic identification of responses suitable for hearing screening. IEEE/EMBS Special Topic Conference on Point-ofCare (POC) Healthcare Technologies: Synergy Towards Better Global Healthcare, January 16–18 2013, Bangalore, India. PHT 2013; art. no. 6461314, 180–183. [32] Valderrama JT, de la Torre A, Alvarez I, Segura JC, Sainz M, Vargas JL. Auditory middle latency responses recorded at high stimulation rates using randomized stimulation and averaging. XXIIIrd International Evoked Response Audiometry Study Group (IERASG) Symposium, June 9–13 2013, New Orleans, LA: 56. [33] Valderrama JT, de la Torre A, Alvarez I, Segura JC, Sainz M, Vargas JL. Deconvolution of overlapping responses and frequency domain-based artifact rejection methods using randomized stimulation. XXIIIrd International Evoked Response Audiometry Study Group (IERASG) Symposium, June 9–13 2013; New Orleans, LA: 57. [34] Valderrama JT, de la Torre A, Alvarez I, et al. Automatic quality assessment and peak identification of auditory brainstem responses with fitted parametric peaks. Comput Meth Prog Bio 2014; 114: 262–275. [35] Valderrama JT, de la Torre A, Alvarez I, et al. A study of adaptation mechanisms based on ABR recorded at high stimulation rate. Clin Neurophysiol 2014; 125: 805–813. [36] Valderrama JT, de la Torre A, Alvarez I, et al. A more efficient use of the recording time with randomized stimulation and averaging (RSA) in hearing screening applications. 9th International Conference of the Saudi Society of Otorhinolaryngology – Head and Neck Surgery, March 4–6 2014, Riyadh, Kingdom of Saudi Arabia. [37] Valderrama JT, Morales JM, Alvarez I, et al. Automatic quality assessment and response detection of auditory evoked potentials based on response tracking. XXIIIrd International Evoked Response Audiometry Study Group (IERASG) Symposium, June 9–13 2013, New Orleans, LA: 55. [38] Wong PKH, Bickford RG. Brain stem auditory evoked potentials: the use of noise estimate. Electroencephalogr Clin Neurophysiol 1980; 50: 25–34.


Supplementary Material Example of MATLAB routine that implements the recording of AER using the conventional method %% PARAMETERS INITIALIZATION fs = 25e3; Name_File = 'EEG_Example'; ER = 1; if(ER) window = 12e-3; Low_freq = 100; High_freq = 3000; Phase_delay = 15; else window = 100e-3; Low_freq = 10; High_freq = 3000; Phase_delay = 3; end AER = zeros(window*fs,1); ISI = 0.030; N_Sweeps = 2000; Click_Duration = 120e-6; Ga = 1250; Gs = 1.0461; Filter_Order = 4; V_ref = 9.8465e-5; I = 70; clear ER window

% Sampling rate % Name of the file % Evoked response: ER=0 for ABR, ER=1 for MLR % % % %

Time window of 12 ms for ABR Low-pass frequency for digital filter High-pass frequency for digital filter Phase distortion compensation (560 us)

% % % %

Time window of 100 ms for MLR Low-pass frequency for digital filter High-pass frequency for digital filter Phase distortion compensation (80 us)

% % % % % % % % %

AER initialization Interstimulus interval of the sequence in ms Number of recorded sweeps Duration of the click in s Gain of the amplifier (calib) Gain of the AD/DA soundcard (calib) Order of the digital filters Absolute intensity level for 0 dBnHL (calib) Intensity level in dBnHL

%% STIMULATION SIGNAL GENERATION x(1:Click_Duration*fs,1) = -1; % Pattern of the rarefaction click h(1:ISI*fs:N_Sweeps*ISI*fs) = 1; % h=1 -> start of the stimuli Seq = conv(x,h); % Signal sequence generation % Channel 1 - Stimulation signal. Channel 2 - Synchronization signal Seq(:,2) = Seq(:,1); % 2-channels sequence t_blocking = floor(length(Seq)/fs); % Recording test time Seq(:,1) = Seq(:,1)*V_ref*10^(I/20); % Seq - intensity level calibrated clear Click_Duration N_Sweeps ISI x h V_ref I %% STIMULATION & RECORDING x = audioplayer(Seq,fs,16); play(x); sound(Seq,fs,16); recorder = audiorecorder(fs,16,2); recordblocking(recorder,t_blocking); y = getaudiodata(recorder); clear t_blocking Seq x recorder %% SCALING EEG = y(:,1)-mean(y(:,1)); EEG = EEG/Ga/Gs*1e6; Sinc = y(:,2)-mean(y(:,2)); clear y Ga Gs

% Remove the offset of the input signal % EEG calibrated in microvolts % Remove the offset of the input signal

%% AER ENHANCEMENT [b a] = butter(Filter_Order,[Low_freq High_freq]*2/fs,'bandpass'); Resp = filter(b,a,EEG); % EEG after digital filtering clear a b Filter_Order Low_freq High_freq

%% SYNCHRONIZATION % Sinc is replaced with samples of amplitude over the 70% of the maximum Sinc = find(Sinc>0.7*max(Sinc)); % Only the first sample is relevant. The following 10 samples are removed. m(1) = Sinc(1); % m(j) - Synchronization samples j = 1; for i=2:size(Sinc,1)-10 if((Sinc(i)-m(j))>10) j = j+1; m(j) = Sinc(i); end end NN = length(m); % NN is the number of recorded sweeps clear Sinc i j %% AER CALCULATION for i=1:NN AER = AER + Resp(m(i):m(i)+length(AER)-1)/NN; % Sweeps averaging end AER = AER(Phase_delay:length(AER)); % Phase distortion compensation clear i %% STORAGE save(Name_File,'AER','EEG','m','NN','fs'); fprintf('Data in \n',Name_File);

AIP/123-QED Auditory brainstem and middle latency responses recorded at fast rates with

VI E

Joaquin T. Valderrama,a) Angel de la Torre, Isaac Alvarez, and Jose Carlos Segura Department of Signal Theory, Telematics and Communications,

RE

CITIC-UGR,

Spain

ND

A. Roger D. Thornton

ER

University of Granada, Granada 18071,

W

randomized stimulation

-U

MRC Institute of Hearing Research, Royal South Hants Hospital, Southampton SO14 OYG,

IA

L

United Kingdom

Manuel Sainzb) and Jose Luis Vargas

ENT Service, Granada 18012, Spain (Dated: August 6, 2014)

CO

NF

ID E

NT

San Cecilio University Hospital,

Fast rate responses with randomized stim. 1

Abstract Randomized stimulation and averaging (RSA) allows auditory evoked This

W

potentials (AEPs) to be recorded at high stimulation rates.

interference derived from overlapping transient evoked responses. This paper analyzes the effects of this interference on auditory brainstem re-

RE

sponses (ABRs) and middle latency responses (MLRs) recorded at rates

VI E

method does not perform deconvolution and must therefore deal with

of up to 300 Hz and 125 Hz, respectively, with randomized stimulation

ER

sequences of a jitter both greater and shorter than the dominant pe-

riod of the ABR/MLR components. Additionally, this paper presents

ND

an advanced approach for RSA (I-RSA), which includes the removal of the interference associated with overlapping responses through an

-U

iterative process in the time domain. Experimental results show that (a) RSA can be efficiently used in the recording of AEPs when the

L

jitter of the stimulation sequence is greater than the dominant period

IA

of the AEP components, and (b) I-RSA maintains all the advantages of RSA and is not constrained by the restriction of a minimum jitter.

NT

The significance of the results of this study is discussed.

ID E

PACS numbers: 43.66.Yw, 43.64.Ri, 43.64.Yp Keywords: iterative-RSA (I-RSA); overlapping transient evoked responses;

CO

NF

high stimulation rate; jitter.

2

I. INTRODUCTION

W

Auditory evoked potentials (AEPs) are a set of low-amplitude voltage peaks (usually less

VI E

than 1 µV at the electrodes), generated in different parts of the auditory pathway in response to a stimulus. AEPs can be classified according to their generator site and their peak latency

(time between the stimulus onset and the occurrence of the peaks), which ranges between

RE

1 ms and 0.5 s. The recording of AEPs is extensively used in both human and animal studies

because of its noninvasive nature. The auditory brainstem response (ABR) is a particular

ER

AEP partly generated in the cochlea, in the auditory nerve and in the brainstem (Eggermont, 2007; Pratt, 2011). The ABR comprises a number of waves that occur within the first 10 ms

ND

from stimulus onset. These waves are identified by sequential Roman numerals, as originally proposed by Jewett and Williston (1971). Although up to seven waves can be identified in

-U

the ABR, the most robust peaks are waves I, III, and V. The recording of ABR signals is commonly used in hospital and clinics worldwide as a hearing screening tool, to detect the

L

hearing threshold, and to detect hearing impairments such as vestibular schwannomas and

IA

Ménière’s disease (Kacker and Deka, 1986; Podoshin et al., 1986; Hall, 2007; Bush et al., 2008). The middle latency response (MLR) is generated in the auditory thalamocortical

NT

system. The MLR has latencies from 10 to 60 ms, and comprises the components Na , Pa , Nb , and Pb (Eggermont, 2007; Pratt, 2011). The longer latency component of the MLR is

ID E

usually affected by attention and is difficult to record under sleep and sedation, limiting the clinical utility of these signals to the assessment of cooperative children and adults (Pratt,

NF

2007). MLR signals are typically used in clinical practice to evaluate the central auditory nervous system and in the assessment of auditory-pathway integrity in cochlear implant

CO

candidates, since electrically elicited MLR signals are less contaminated by the stimulus artifact than electrical ABR signals (Fifer and Sierra-Irizarry, 1988; Hall, 2007; Pratt, 2007, a)

Author to whom correspondence should be addressed. Electronic mail: jvalderrama@

ugr.es b) Also at Dept. of Surgery and its Specialties, University of Granada, Granada 18012, Spain. 3

2011). AEPs are conventionally elicited by stimuli presented periodically (Wong and Bickford,

W

1980; Elberling and Don, 2007), i.e., with a constant inter-stimulus interval (ISI). This

VI E

method has the limitation that the ISI must be greater than the averaging window to

avoid contamination of the recording by the adjacent response; otherwise it would not be

RE

mathematically possible to recover the overlapping AEP (Zollner et al., 1976; Kjaer, 1980; Jewett et al., 2004). Considering standard averaging windows of 10 ms for ABR signals,

and 100 ms for MLR signals, ABRs and MLRs cannot be recorded with the conventional

ER

method at rates higher than 100 Hz and 10 Hz, respectively. However, the recording of AEPs at higher rates presents certain advantages, as several authors have reported. First,

ND

the recording of AEPs at high rates allows the study of neural adaptation (Lasky, 1997; Burkard et al., 1990; Valderrama et al., 2014a). Other authors state that high stimulation

-U

rates may improve accuracy in estimating the hearing threshold of a subject (Leung et al., 1998). High stimulation rates have also been used to detect certain pathologies (e.g., Don

L

et al., 1977; Stockard et al., 1978; Yagi and Kaga, 1979; Jiang et al., 2000; Thornton et al.,

IA

2006; Bohorquez et al., 2009). Additionally, some authors have concluded that the use

NT

of high stimulation rates may speed up hearing screening, since less recording time would be necessary in order for a specific number of averaged auditory responses (sweeps) to be

ID E

obtained (Thornton and Slaven, 1993; Leung et al., 1998; Bell et al., 2001, 2002). However, neural adaptation produces changes in the morphology of the responses, decreasing the signal-to-noise ratio (SNR) of the response. Whether or not high rates are useful in the

NF

recording of AEPs in less time is currently controversial (Burkard and Don, 2007). Various methods have emerged to overcome the rate limitation imposed by the conven-

CO

tional technique. These methods use jittered stimulation sequences with specific properties in the time and frequency domains that allow the recovery of overlapping transient evoked responses. The jitter of a stimulation sequence measures the dispersion of the ISI compared with a periodical presentation of stimuli, with constant ISI. The most relevant methods used to obtain AEPs at high rates are maximum length sequences (MLS) (Eysholdt and Schreiner, 4

1982), quasiperiodic sequence deconvolution (QSD) (Jewett et al., 2004), continuous loop averaging deconvolution (CLAD) (Ozdamar et al., 2003a,b; Delgado and Ozdamar, 2004;

W

Ozdamar and Bohorquez, 2006), least-squares deconvolution (LS) (Bardy et al., 2014a), and

VI E

randomized stimulation and averaging (RSA) (Valderrama et al., 2012). The MLS, QSD, CLAD, and LS methods obtain the AEP by a deconvolution procedure. The fundamentals of deconvolution of overlapping responses are described below. The recorded EEG y(t)

RE

can be represented as the convolution of a stimulation signal s(t) and an AEP h(t) plus

noise n(t): y(t) = s(t) ∗ h(t) + n(t). In the frequency domain, this equation would yield

ER

Y (f ) = S(f ) · H(f ) + N (f ). The AEP in the frequency domain H(f ) can be worked out as H(f ) = Y (f )/S(f ) − N (f )/S(f ). In this equation, the value of N (f ) is unknown. There-

ND

b ) = Y (f )/S(f ), considering fore, the AEP in the frequency domain can be estimated as H(f

N (f )/S(f ) as the error between the estimated and the real AEP. Accurate estimation of

-U

b ) requires (a) the reduction of the power of the noise distribution N (f ) by averaging, and H(f (b) the selection of a stimulation sequence s(t) whose frequency components S(f ) are not

L

close to zero, otherwise the noise at that frequency would be amplified. Finally, estimation of

IA

b ): b b )} the AEP in the time domain is the inverse Fourier transform of H(f h(t) = IF F T {H(f (Jewett et al., 2004; Ozdamar and Bohorquez, 2006; Valderrama et al., 2014a). The MLS

NT

method has been widely used not only with AEPs (e.g., Leung et al., 1998; Bohorquez and Ozdamar, 2006; Lavoie et al., 2010), but also with transient evoked otoacoustic emissions

ID E

(OAE) (e.g., Hine et al., 2001; Thornton and Slaven, 1993; de Boer et al., 2007). The distribution of the ISI in this method is adjusted to De-Bruijn sequences, in which a k-ary de-Bruijn

NF

sequence B(k, n) of order n of a given alphabet A is a pseudorandom cyclic sequence with size k for which every possible subsequence of length n appears in the sequence exactly once

CO

(Tuliani, 2001; Burkard et al., 1990). The nature of De-Bruijn sequences imposes on MLS the restriction of a very high jitter (Jewett et al., 2004; Ozdamar et al., 2007). Some authors believe that high-jittered stimulation sequences are a disadvantage in recording AEPs, because the morphology of the response associated with a stimulus not only depends on the averaged stimulation rate, but also on the preceding ISI; therefore, high-jittered stimulation 5

sequences could evoke auditory responses of different morphology (especially at high rates), and the assumption of a time-invariant response would not be accomplished (Jewett et al.,

W

2004; Ozdamar and Bohorquez, 2006; Valderrama et al., 2014a). The QSD, CLAD, and LS

VI E

methods present different approaches to deconvolve overlapping auditory responses evoked by low-jittered stimuli. The methods based on deconvolution have been efficiently used in

Presacco et al., 2010; Wang et al., 2013; Bardy et al., 2014b).

RE

several clinical and research applications (e.g., Bohorquez et al., 2009; Gutschalk et al., 2009;

In contrast to the MLS, QSD, CLAD, and LS methods, RSA does not perform de-

ER

convolution. The approach taken by the RSA method consists of averaging a number of sweeps, corresponding to a burst of stimuli in which the ISI varies randomly according to a

ND

predefined probability distribution (randomized stimulation). The RSA method includes a digital blanking process, which considers null values in the average process any samples of

-U

the EEG contaminated by stimulus artifact. The main advantages of the RSA method are that (a) it facilitates precise control of the jitter of the stimulation sequence, (b) random-

L

ized stimulation sequences are easy to generate, since they are not subject to restrictions

IA

in the frequency domain, because RSA does not perform deconvolution, and (c) it allows

NT

sweeps to be processed separately. These particular advantages were used in Valderrama et al. (2014a) to carry out a study of the fast and slow mechanisms of neural adaptation

ID E

in humans through the separated responses method, which is based on the categorization of sweeps according to their preceding ISI. The categorization of responses according to their preceding ISI may be accomplished because of the individual processing of sweeps that is

NF

allowed by RSA. Additionally, the separate processing of responses allows artifact-rejection techniques to be used more efficiently. In RSA, each sweep can be individually accepted or

CO

rejected for averaging. In contrast, deconvolution-based methods process blocks of responses, resulting in a less flexible application of the artifact rejection procedure, since the portions rejected for averaging in these methods are greater than in RSA. The RSA method does, however, have some limitations. Since RSA does not perform deconvolution, this method must deal with interference derived from the contamination produced by overlapping ad6

jacent responses. Interference associated with overlapping responses can be reduced with averaging in RSA provided that the amount of jitter is large enough to enable positive and

W

negative components of such interference to be cancelled with averaging (Valderrama et al.,

VI E

2012). In addition to this, the digital blanking process in RSA entails non-uniform averaging of auditory responses along the averaging window, and small amounts of jitter could lead

to significant differences in terms of quality between different segments of the response. For

RE

these reasons, the RSA method requires stimulation sequences with a minimum amount of jitter.

ER

In this paper we analyze the performance of the RSA method in different jittering conditions, and we present a new approach for RSA that includes the estimation and suppression of

ND

the interference associated with overlapping adjacent responses through an iterative process in the time domain. We have called this evolution of RSA “iterative-randomized stimulation

-U

and averaging” (I-RSA). This new approach maintains the advantages of RSA while eliminating the need for a digital blanking process, and thus also eliminating the limitation of the

L

minimum amount of jitter that is imposed in RSA. In this paper, we analyze the interference

IA

associated with overlapping responses in the RSA and I-RSA methods with both real and artificially synthesized ABR and MLR signals obtained at different stimulation rates with

NT

both long and short jitter distributions. Portions of this research were presented at the International Evoked Response Audiometry Study Group (IERASG) 2013 meeting held in

ID E

New Orleans, LA (Valderrama et al., 2013).

NF

II. METHODS

CO

A. Randomized stimulation and averaging The randomized stimulation and averaging (RSA) method consists of averaging auditory

responses, corresponding to a burst of stimuli in which the inter-stimulus interval (ISI) varies randomly according to a predefined probability distribution (randomized stimulation). The RSA technique involves a digital blanking process to minimize the effect of the stimulus 7

artifact in overlapping responses. The digital blanking process considers as null values any EEG samples in which stimulus artifact occurs. Using RSA notation (Valderrama et al.,

W

2012), let y(n), s(n) (n = 1, . . . , N ), J, and N be, respectively, the digitized EEG, the

VI E

synchronization signal (indicating with the value of 1 the start of each stimulus, and 0 otherwise), the length of the averaging window, and the total number of EEG samples.

RE

Considering a stimulation sequence with K stimuli, the index of the samples in which each stimulus starts can be represented by m(k) (k = 1, . . . , K). Hence, s m(k) = 1 ∀k. The blanking signal b(n) differentiates valid samples of the EEG (value 1) from samples

ER

contaminated by the stimulus artifact (value 0). The duration of blanking depends on the duration of stimulus artifact. The implementation of blanking of 1 ms duration is appropriate

ND

for short duration stimuli, e.g., clicks. In this case, the blanking signal considers as null values 0.2 ms before and 0.8 ms after each stimulus (equation 1). Longer-duration stimulus

-U

artifacts would require a longer-duration blanking signal. The AEP x b(j) is estimated in RSA

by averaging the sections of the digitized EEG not contaminated by the stimulus artifact

NT

0 if n [m(k) − 0.2ms · fs , m(k) + 0.8ms · fs ], ∀ k

NF

n = 1...N

(1)

j = 1...J

(2)

1 otherwise

ID E

b(n) =

(

IA

L

(equation 2).

x b(j) =

K X k=1

b(m(k) + j) · y(m(k) + j) K X

b(m(k) + j)

k=1

The RSA method is fully described in Valderrama et al. (2012). In comparison with

CO

other methods based on deconvolution, RSA (a) allows a precise control of the jitter of the stimulation sequence, (b) presents generation of stimulation sequences that is not subject to any particular constraint in the frequency domain, and (c) allows auditory responses to be processed separately. However, since RSA does not perform deconvolution, this method must deal with interference associated with overlapping adjacent responses, and the amount 8

of jitter of the stimulation sequences in RSA must therefore be greater than the dominant period of the AEP components in order for positive and negative components of this inter-

W

ference to be cancelled with averaging. The dominant period of an AEP can be estimated

VI E

through the autocorrelation function as the shift at which the closest maximum correlation occurs (Oppenheim and Schafer, 1999). Figure 2 shows the autocorrelation function for

given high-quality ABR and MLR signals. This figure shows that the dominant period of

RE

ABR and MLR signals is about 2 ms and 25 ms, respectively, which is consistent with previ-

ous studies (Rudell, 1987; Delgado and Ozdamar, 1994; Galambos et al., 1981; Picton et al.,

ER

1992; Pastor et al., 2002; Pratt, 2007). Moreover, the digital blanking process included in RSA causes non-uniform averaging along the averaging window, i.e., the number of aver-

ND

aged samples along the averaging window varies according to the amount of jitter of the stimulation sequence and the duration of digital blanking. Averaging a number of auditory

-U

responses below a threshold of 70% of the available samples could produce noticeable differences in terms of quality between different segments of the response. Averaging at least

L

70% of the available samples ensures differences in quality between different segments of the

IA

response that are less than 1.55 dB. Other averaging thresholds, e.g., 50%, 25%, or 10%, would produce a maximum quality loss of, respectively, 3 dB, 6 dB, and 10 dB. Figure 1

NT

shows an analysis of the influence of the amount of jitter (upper panel) and the duration of the digital blanking (lower panel) on the number of averaged samples along the averaging

ID E

window of an ABR signal (10 ms) with the RSA method. The stimulation sequences of this study were generated using 20 000 stimuli randomly distributed with a uniform distribution

NF

of probability between ‘a’ and ‘b’ ms (ISIa−b ). Analysis of the upper panel shows the influence of the amount of jitter on the number of averaged samples when the duration of digital

CO

blanking is 1 ms. In this analysis, using a 4 ms jittered stimulation sequence (ISI5−9 ), the

lowest number of averaged samples is around 15 000 samples; in a 2 ms jittered sequence (ISI5−7 ), that number is about 10 000 samples; and in a 0.5 ms jittered sequence (ISI5−5.5 ),

there would be a segment in the averaging window that could not be obtained. Analysis of the lower panel shows the influence of long- and short-duration blanking on the number 9

of averaged samples along the averaging window. This analysis shows that long duration blanking used in long duration stimuli would impose the restriction of a large amount of

W

jitter to meet the requirement of averaging at least 70% of the available samples. The use

VI E

of a shorter-duration blanking could allow less jittered stimulation sequences to be implemented. This study shows that the amount of jitter and the duration of digital blanking

are parameters influencing the number of averaged samples along the averaging window.

RE

The RSA method requires (a) a minimum amount of jitter that must be greater than the

dominant period of the components of the AEPs to allow positive and negative components

ER

of the interference associated with overlapping responses to be cancelled by averaging; and (b) a jitter distribution that allows a sufficient number of averaged samples all along the

ND

averaging window to avoid appreciable differences in quality between different segments of the recorded AEP.

-U

With this in view, the authors have developed a modified version of RSA which does not require a digital blanking process and eliminates the interference associated with overlapping

L

adjacent responses through an iterative process in the time domain. We have called this

IA

approach “iterative-randomized stimulation and averaging” (I-RSA). This version of RSA

NT

maintains the main properties of RSA, while overcoming the limitation of the jitter imposed in RSA. The approach for this method is based on iterations that include estimation of

ID E

the interference associated with overlapping responses, its subtraction from the recorded EEG, and re-estimation of the AEP. The improved AEP estimate on each iteration leads to a better estimate of the interference associated with overlapping responses, and a better

NF

AEP estimate can therefore be obtained recursively. The accuracy of the AEP estimate increases with the number of iterations. The total number of iterations can be set either

CO

as a predefined value, or automatically, in which case the method stops iterating when the differences between AEP estimates in successive iterations are negligible. As in RSA, the generation of stimulation sequences in I-RSA is based on randomized stimulation, where the ISI of the stimuli varies randomly according to a predefined probability distribution (Valderrama et al., 2012). 10

The mathematical formulation of I-RSA is described below.

Using RSA notation

(Valderrama et al., 2012), the estimate of the transient evoked potential x b(j) (j = 1, . . . , J)

W

is obtained in I-RSA by an iterative process in the time domain. Each iteration (i) results

VI E

in an estimate of the transient evoked potential, represented by b hi (j) (j = 1, . . . , J). The

AEP estimate in each iteration by the I-RSA method is obtained as the average of the K

K 1 X b hi (j) = yk j + m(k) K k=1

RE

sweeps whose corresponding overlapping responses are suppressed: j = 1, . . . , J

(3)

ER

where yk (n) (n = 1, . . . , N ) represents the EEG in which the overlapping responses associated with the stimulus k are suppressed. The yk (n) signals can be obtained for each stimulus

ND

as the original electroencephalogram y(n) minus the AEP estimates on the preceding iteration (b hi−1 ) corresponding to all stimuli excluding the stimulus k: n = 1, . . . , N

(4)

-U

yk (n) = y(n) − s(n) − sk (n) ∗ hi−1

where sk (n) represents the stimulation signal for the stimulus k, and the symbol ∗ is the

L

convolution operator. Considering the signal σ(n) as the original EEG with all AEPs sup-

IA

pressed: σ(n) = y(n) − s(n) ∗ hi−1 , then:

yk (n) = y(n) − s(n) ∗ hi−1 + sk (n) ∗ hi−1 = σ(n) + sk (n) ∗ hi−1

(5)

NT

n = 1, . . . , N

the sections of yk corresponding to the averaging window of the stimulus k can thus be

ID E

obtained as:

yk j + m(k) = σ j + m(k) + sk j + m(k) ∗ hi−1

j = 1, . . . , J

(6)

NF

The sk (n) signal can be represented as δ n − m(k) , where δ(n) represents the Dirac delta

CO

function, with the value 1 for n = 0, and 0 otherwise. Hence:

yk j + m(k) = σ j + m(k) + δ j + m(k) − m(k) ∗ hi−1

j = 1, . . . , J

(7)

Since the convolution of the delta function with any other tempered distribution S is simply the distribution S: yk j + m(k) = σ j + m(k) + δ(j) ∗ hi−1 = σ j + m(k) + hi−1 11

j = 1, . . . , J

(8)

and therefore, from equation 3, the AEP estimation in iteration i can be obtained as:

K 1 X σ j + m(k) + hi−1 K k=1

In this equation, the parameter

1 K

PK

k=1

j = 1, . . . , J

(9)

VI E

=

W

K K 1 X 1 X b σ j + m(k) + hi−1 = hi (j) = yk j + m(k) = K k=1 K k=1

σ j + m(k) represents the correction made to

RE

the AEP estimate on the preceding iteration (hi−1 ). Under certain jitter distributions, this correction parameter may cause instability problems, leading to worse AEP estimates in

ER

successive iterations. This instability issue can be solved by inserting a correction factor (α), which may constrain (α-values lower than 1) or enhance (α-values lower than 1) the

ND

correction made to hi−1 . Thus, the AEP estimate in iteration i is obtained as: K 1 X b hi (j) = α · σ j + m(k) + hi−1 K k=1

j = 1, . . . , J

(10)

-U

The estimated AEP on each iteration b hi (j) is used in the following iteration as b hi−1 (j).

The I-RSA method is initialized with b h0 (j) = 0 ∀j. Finally, the estimate of the transient

L

AEP by I-RSA (b x(j)) can be obtained either as the estimated AEP after a predefined number

IA

of iterations (I), or when the differences between the AEP estimates in successive iterations

or

ID E

x b(j) = b hI (j)

NT

are negligible:

x b(j) = b hi (j) / b hi (j) ≈ b hi−1 (j)

j = 1, . . . , J

(11)

The speed of convergence of this method towards the AEP estimate is related to the α-value and to the level of interference associated with overlapping responses, which depends

NF

on (a) the amount of jitter of the stimulation sequence and (b) the averaged ISI compared with the length of the averaging window. For a given jitter distribution, low α-values may

CO

slow down the speed of convergence of this method toward the AEP estimate, but nonetheless assure convergence. However, when instability is not a problem, high α-values may speed-up convergence toward the AEP. Figure 3 illustrates an example of an iteration of the I-RSA method outlined above. In this example, the sampling frequency is fS = 25 kHz, the length of the averaging window 12

is J = 2500 samples, which corresponds to 100 ms, and the correction factor is α = 1. Figure 3.A shows an MLR signal as an example of transient AEP. The ABR, Na , Pa , Nb ,

W

and Pb components can be identified in this transient AEP. Figure 3.B shows a segment of

VI E

a synchronization signal (s(n)) whose interstimulus intervals vary randomly with a uniform distribution between 42 and 58 ms (ISI42−58 ). The index of the samples at which stimuli

start (m(k), k = 1 . . . , 5) are shown next to each stimulus. This figure also shows the

RE

recorded electroencephalogram (EEG), which in this example was artificially synthesized by

the convolution of the synchronization signal and the transient AEP plus white noise added

ER

to the EEG at a signal-to-noise ratio of 9.5 dB. Figure 3.C shows the estimated AEP on the preceding iteration: hi−1 (j), which is used in the iteration i as an approximation of the AEP

ND

to evaluate and suppress the interference associated with overlapping responses. Figure 3.D shows the recorded EEG with all overlapping responses subtracted (σ(n)). Figure 3.E out-

-U

lines the AEP-estimation process for the iteration i, which is obtained by adding the AEP estimation in the preceding iteration hi−1 (j) to the average of the signals σi · · · σK .

L

The performance of the RSA and I-RSA methods is assessed in this paper with both real

IA

and artificially synthesized ABR and MLR signals using stimulation sequences of different

NT

rates and jitter distributions.

ID E

B. EEG recording

The AEP-recording procedure consisted of the presentation of auditory stimuli to the

NF

subjects and the recording of their associated electrical responses (sweeps). Stimulation of the auditory system was performed monaurally by monophasic 0.1 ms rarefaction clicks to

CO

evoke a synchronous firing of neurons, especially those in the 1000−4000 Hz region (Hall,

2007; Thornton, 2007), using an insert earphone (ER-3A Etymotic Research, Elk Grove Village, IL). The recording sessions took place in a shielded screening booth prepared to attenuate acoustical and electromagnetic interference. The subjects were seated comfortably in order to minimize electromyogenic interference. The intensity level 0 dB normal-hearing 13

level (nHL) was established as the averaged threshold level (intensity level at which stimuli are just detectable) of a group of 15 subjects (9 male and 6 female) aged 24-31, with

W

no self-reported history of auditory dysfunction. This intensity level corresponds to 33.54

VI E

dB peak-to-peak equivalent sound pressure level (dB peSPL), calibrated by an Artificial Ear Type 4153 2-cc acoustic coupler (Br¨ uel & Kjær Sound & Vibration Measurement A/S,

Nærum, Denmark). The EEGs were recorded by Ag/AgCl surface electrodes placed on the

RE

upper forehead (active), lower forehead (ground), and ipsilateral mastoid (reference). The

interelectrode impedance was below 5 kΩ in all recordings. The EEG was amplified by 70 dB

ER

and bandpass filtered by a 24 dB/Octave slope filter with a bandwidth of [150−3500] Hz for ABR and [0.5−3500] Hz for MLR. The recorded signal was sampled at 25 kHz, digitally

ND

filtered (4th order, bandwidth: [300−3000] Hz for ABR, [30−1000] Hz for MLR), and stored using 16 bits of quantization. Digital signals were processed with algorithms implemented

-U

in Matlab (The Mathworks, Inc., Natick, MA). An artifact-rejection method prevented the processing of any sweeps whose maximum amplitude exceeded ± 10 µV. No recordings were

L

processed from the first second of the recording test in order to acclimatize the subject to

IA

the stimulation sequence and prevent the processing of any event-related potentials (ERPs) associated with novelty. A fuller description of the recording system used in this study

NT

can be found in Valderrama et al. (2011, 2014c). The recording process carried out in this study was in accordance with the Code of Ethics of the World Medical Association (Decla-

ID E

ration of Helsinki) for experiments involving humans, and it was approved by the Research Ethics Committee established by the Health Research Authority (Reference number RHM

CO

NF

ENT0082).

C. Description of the experiments This study involves the recording of ABR and MLR signals at different stimulation rates

using randomized stimulation and averaging (RSA) and iterative-randomized stimulation and averaging (I-RSA) methods. Sixteen subjects, 10 males and 6 females, aged 19−46 14

(with a mean and standard deviation age of 30 ± 6) were recruited for this study. No participant showed any significant auditory dysfunction, presenting audiometric thresholds

W

of 20 dB HL or less for pure tones between 250 and 8000 Hz. The subjects were volunteers

VI E

and were informed in detail about the experimental protocol and possible side effects of the test. A consent form was signed by the participants before the beginning of the recording

session, which was carried out at the Royal South Hants Hospital (Southampton, United

RE

Kingdom). Subjects #S1 to #S8 participated in the study of ABR signals, and subjects #S9 to #S16 participated in the study of MLR signals. The RSA and I-RSA methods were

ER

implemented as presented in section II.A of this paper. In RSA, the digital blanking used in the RSA method was 1 ms ([-0.2 to 0.8] ms). In I-RSA, the total number of iterations was

ND

I = 20, and the correction factor was α = 0.8. The randomized stimulation sequences used in this study were built according to uniform jitter distributions, i.e., the ISI of an ISIa−b

-U

stimulation sequence varies randomly with uniform distribution between ‘a’ and ‘b’ ms (Valderrama et al., 2012).

L

The first experiment consists of a study of the interference associated with overlapping

IA

responses in both real and computer simulated ABR and MLR signals obtained at different

NT

stimulation rates with the RSA and the I-RSA methods in different jittering conditions. The aim of this study is to assess the performance of the proposed method (I-RSA), and

ID E

to analyze the effects of the interference associated with overlapping responses with ABR and MLR signals obtained from the RSA and I-RSA methods when the amount of jitter is greater than, equal to, and shorter than the dominant period of the ABR/MLR compo-

NF

nents. The study of ABR signals included the generation of stimulation sequences at rates of 71 Hz (8500 stimuli), 83 Hz (10 000 stimuli), 100 Hz (12 000 stimuli), 125 Hz (15 000 stim-

CO

uli), 167 Hz (20 000 stimuli), 250 Hz (30 000 stimuli) and 300 Hz (36 000 stimuli), using jitter distributions greater than (4 ms), equal to (2 ms), and shorter than (0.6 ms) the dominant period of the ABR components (about 2 ms). The recording of MLR signals was performed by generating stimulation sequences at rates of 8 Hz (2000 stimuli), 20 Hz (2400 stimuli), 40 Hz (4800 stimuli), 67 Hz (8000 stimuli), 100 Hz (12 000 stimuli), and 15

125 Hz (15 000 stimuli), using jitter distributions greater than (50 ms), equal to (25 ms), and shorter than (16 ms) the dominant period of the MLR components (about 25 ms). The

W

varying number of stimuli used in these stimulation sequences was set to achieve a compro-

VI E

mise between the duration of the test and the quality of the response. An averaging window of 25 ms (J = 625 samples) was used in the study of ABR signals to expand the analysis

of the interference associated with overlapping adjacent responses outside the standard av-

RE

eraging window (10 ms). The length of the averaging window for MLR signals was 100 ms

(J = 2500 samples). In the study with computer-simulated signals, real high-quality ABR

ER

and MLR signals were used as templates. The template used for the ABR test was recorded from subject #S8 (male, aged 26), using 20 000 stimuli presented at 70 dB nHL with a

ND

stimulation sequence ISI20−24 (45 Hz). The template used for the MLR test was recorded from subject #S9 (male, aged 28), with stimuli presented at 70 dB nHL using a stimulation

-U

sequence ISI117−133 (8 Hz) of 10 000 stimuli. Different EEGs were artificially synthesized by the convolution of the templates with each stimulation sequence of this study. An estimate

L

of the template was obtained by the RSA and I-RSA methods from the EEG synthesized

IA

at each stimulation rate and under each jittering condition. These artificially synthesized EEGs do not include any noise artifacts typically present in the recording of real AEPs, such

NT

as interference derived from the myogenic activity of the subject or from electromagnetic interference. The only type of interference included in these synthesized EEGs is associated

ID E

with overlapping responses, which was estimated as the root mean square (RMS) value of the difference between the template and the signals obtained by each method. In the study

NF

with real signals, the ABR and MLR signals were recorded from subject #S9 (male, aged 28) using stimuli presented at 70 dB nHL. The real ABR and MLR signals were obtained

CO

from these recorded EEGs using the RSA and I-RSA methods. The second experiment includes an analysis of latencies and amplitudes of the peaks of

real ABR and MLR signals obtained with RSA and I-RSA at different stimulation rates in a set of 8 normal-hearing subjects for the study of ABR signals (6 males and 2 females, aged 16−36), and a different set of 8 normal-hearing subjects for the study of MLR signals (4 males 16

and 4 females, aged 23−46). ABR signals were elicited by stimuli presented at 70 dB nHL at rates of 45, 56, 71, 83, 100, 125, and 250 Hz (20 000 stimuli in all rates), using jitter

W

distributions of 4 ms (greater than the dominant period of the ABR components). MLR

VI E

signals were elicited by stimuli presented at 70 dB nHL at the rates 8 Hz (2000 stimuli),

20 Hz (5000 stimuli), 40 Hz (5000 stimuli), 67 Hz (5000 stimuli), 100 Hz (10 000 stimuli), and 125 Hz (20 000 stimuli), using jitter distributions of 16 ms (shorter than the dominant period

RE

of the MLR components). The varying number of stimuli used in the stimulation sequences allowed MLR signals of sufficient quality to be recorded with an appropriate recording time.

ER

The length of the averaging window was 10 ms for ABR signals (J = 250 samples), and 100 ms for MLR signals (J = 2500 samples). Latencies (Lpeak ) were measured as the

ND

difference in milliseconds between the stimulus onset and the maximum value of the peak. In ABR signals, amplitudes (Apeak ) were measured in microvolts as the difference between

-U

the top of the peak and the following trough; whereas in MLR signals, amplitudes were measured as the difference between the negative and the positive wave complex (Hall, 2007;

L

Burkard and Don, 2007; Pratt, 2007). The differences between the morphology of the AEPs

IA

obtained with the RSA and I-RSA methods were analyzed by a matched paired t-test for differences in latencies (LRSA −LI−RSA ) and by a matched paired Wilcoxon signed rank test

NT

for the ratio of amplitudes calculated as ARSA /AI−RSA − 1, using a significance level of

ID E

α = 0.05 in both analyses.

NF

III. RESULTS

CO

A. Experiment 1 Figure 4 shows real and computer-simulated ABR and MLR signals obtained with the

RSA and I-RSA methods at different rates using stimulation sequences of jitter distributions greater than, equal to, and shorter than the dominant period of the ABR/MLR components. In ABR signals, the averaging window of 25 ms allows the interference associated with overlapping adjacent responses to be studied outside the standard averaging window for 17

ABR signals (10 ms). In the computer-simulated study, the signals used as a template are shown below each figure. The main components of these AEPs are labeled on the figure.

W

The level of interference associated with overlapping responses has been estimated in each

VI E

method as the RMS value of the difference between the template and the obtained signals

by the RSA and I-RSA methods. The study with simulated signals shows that estimation of the ABR and MLR signals by the RSA method is accurate when the distribution of the

RE

jitter is greater than the dominant period of the ABR/MLR components. The ABR signals

corresponding to a jitter distribution of 4 ms and the MLR signals corresponding to a jitter

ER

distribution of 50 ms present a similar morphology to their corresponding template (with similar latencies and amplitudes of their components). When the jitter distribution is equal

ND

to the dominant period of the ABR/MLR components, i.e., jitter of 2 ms for ABRs and 25 ms for MLRs, the ABR and MLR signals estimated by the RSA method present small

-U

differences with the template waveform owing to the interference associated with overlapping responses. For instance, some additional peaks appear outside the averaging window in

L

ABR signals, and the components wave I at 125 Hz; wave II at 167 Hz; waves I, III, and

IA

VII at 250 Hz; and the Pb at 20 Hz are slightly overestimated. The effects of interference associated with overlapping responses become particularly manifest when the jitter is shorter

NT

than the dominant period of the ABR/MLR components, e.g., jitter of 0.6 ms for ABRs and 16 ms for MLRs. At rates up to 100 Hz, the ABR signals estimated by RSA perfectly

ID E

fit the template waveform, although a discernible waveform similar in morphology to an ABR signal appears next to the response (between 10 and 25 ms) owing to the interference

NF

associated with overlapping responses. As the stimulation rate increases, these additional components (which are due to interference, and are not part of the true response) appear on

CO

the latencies of ABR components (first 10 ms) and produce contamination of the response, which cannot be reduced by averaging. Some of these effects are overestimation of waves I and VII at 125 Hz; waves III, VI, and VII at 167 Hz; waves I, V, and VII at 250 Hz, and wave II at 300 Hz. Additionally, this figure shows that the ABR signals obtained for a jitter of 0.6 ms present noticeable differences in quality between different segments of 18

the response, and certain segments could not be obtained as a result of digital blanking. Analysis of the effects of overlapping responses in MLR signals shows significant differences

W

between the MLRs obtained with RSA and the template. At 20 Hz, these effects cause

VI E

overestimation of the Pb component and the generation of an additional peak at about

80 ms. At 40 Hz, the effects of overlapping responses cause significant overestimation of the Na , Pa , Nb , and Pb components, and an additional peak is also generated at latency of about

RE

80 ms. These effects are a consequence of the resonance generated when the components are in phase (occurring at the same time relative to the stimulus) and overlap (Bohorquez

ER

and Ozdamar, 2008). This phenomenon is generally known as 40-Hz event-related potential (ERP) and was first described by Galambos et al. (1981). The latencies of the components

ND

are estimated correctly at rates of 20 Hz and 40 Hz. At rates of 67 Hz, 100 Hz, and 125 Hz the interference associated with overlapping responses causes underestimation of the amplitudes

-U

of all components. Underestimation of the amplitude of the Pb component at these rates leads to premature estimation of its latency. In contrast to RSA, the computer-simulated

L

study shows that the I-RSA method is able to estimate the true AEP (template) accurately

IA

under all recording conditions. The ABR and MLR signals obtained by the I-RSA method present the same morphology as the template signal, and the interference associated with

NT

overlapping responses in I-RSA is lower than 0.01 µVRM S for all ABR and MLR signals at

ID E

all stimulation rates and under all jittering conditions. Analysis of the morphology of real ABR and MLR signals obtained with the RSA and I-RSA methods is consistent with the computer-simulated study. This study shows no

NF

significant differences between the ABR and MLR signals obtained with the RSA and I-RSA methods when the jitter of the stimulation sequences is greater than (4 ms for ABRs and

CO

50 ms for MLRs) or equal to (2 ms for ABRs and 25 ms for MLRs) the dominant period of the ABR/MLR components. The real ABR signals obtained with RSA for a jitter of 0.6 ms at rates of 71, 83, 100, and 125 Hz show additional components similar in morphology to ABR signals appearing with latencies of 10−25 ms. At rates higher than 100 Hz these additional components (which are not part of the response) appear within the first 10 ms of 19

the averaging window and contaminate the response, producing, for example, overestimation of wave I at 125 Hz, of wave II at 167 Hz, and of wave IV at 250 Hz. In contrast, the ABR

W

signals obtained with the I-RSA method present no additional components, and the changes

VI E

in the morphology of the ABR signals across stimulation rates is consistent with previous

literature: as the stimulation rate increases, the amplitude of the components decreases and

the latency increases (to a greater extent the more central the components) as a consequence

RE

of neural adaptation (Lasky, 1997; Burkard et al., 1990; Valderrama et al., 2014a). The real MLR signals obtained with RSA and I-RSA for a jitter of 16 ms show discernible

ER

differences consistent with the computer-simulated analysis. Taking the signals obtained with I-RSA as a reference, the Pb component on the MLR signal obtained with RSA at

ND

20 Hz is overestimated, the Na -Pa and Nb -Pb components are overestimated at 40 Hz, and the Na -Pa and Nb -Pb components are underestimated at rates of 67 Hz, 100 Hz, and 125 Hz.

-U

These results highlight that (a) the performance of the RSA method is appropriate when the jitter of the stimulation sequence is greater than the dominant period of the ABR/MLR

L

components, and (b) the I-RSA method is able to suppress the interference associated to

IA

overlapping responses, leading to accurate estimates of ABR and MLR signals when the

NT

jitter is either greater or shorter than the dominant period of the ABR/MLR components.

ID E

B. Experiment 2

Figure 5 shows ABR and MLR signals recorded on two sets of 8 normal-hearing sub-

NF

jects (16 participants) at different stimulation rates with the RSA and I-RSA methods. Overlapping the AEPs recorded with the RSA and I-RSA methods under each recording

CO

condition allows their differences to be appreciated. This experiment includes an analysis of amplitudes and latencies of ABR and MLR signals obtained by both methods. Figure 6 shows the mean (and standard deviation in errorbars) of the latencies and

amplitudes of the main components of ABR and MLR signals obtained by the RSA and I-RSA methods at different stimulation rates. The values of these parameters estimated 20

in both ABR and MLR signals by the I-RSA method are consistent with other studies reporting AEPs obtained using the MLS and CLAD methods (Lasky, 1984; Lina-Granade

W

et al., 1993; Leung et al., 1998; Stone et al., 2009; Bell et al., 2001, 2002; Ozdamar et al.,

VI E

2007). The large standard deviation of the analysis of amplitudes in this study points

toward significant variability among subjects in terms of amplitudes. With regard to ABR signals, the analysis of latencies and amplitudes of waves I, III, and V indicates that as

RE

the stimulation rate increases: (a) the latency of wave I experiences a slight positive shift, (b) the latency of wave III undergoes a moderate positive shift, (c) the latency of wave

ER

V increases in more deeply, and (d) the amplitude of all the waves decreases. The deeper shift of wave V in comparison with wave III highlights that the stimulation rate influences

ND

central components to a greater extent than peripheral components, which is consistent with previous studies (Pratt and Sohmer, 1976; Yagi and Kaga, 1979; Jiang et al., 2009). The

-U

analysis of latencies and amplitudes of MLR signals obtained with the I-RSA method at different stimulation rates shows that as stimulation rate increases: (a) the latency of the

L

Na and Pa components remains fairly constant, (b) the latency of the Nb and Pb components

IA

decreases, (c) the amplitude of the Na -Pa component decreases, and (d) the amplitude of the Nb -Pb component increases at 40 and 67 Hz and decreases at other rates. The resonance

NT

of the Nb -Pb amplitude at 40 and 67 Hz is possibly due to the association of this component with the auditory primary thalamo-cortical pathway at low rates, and with the non-primary

ID E

reticulo-thalamo-cortical pathway at rates near 50 Hz (Ozdamar et al., 2007). A comparison of the latencies and amplitudes of AEPs obtained by the RSA and I-RSA

NF

methods is shown in tables I and II. These tables show the mean (and standard deviation in parentheses) of the differences of latencies expressed in milliseconds (LRSA −LI−RSA ) and

CO

ratio of amplitudes (estimated as ARSA /AI−RSA − 1) between AEPs obtained by the RSA

and I-RSA methods. Asterisks represent statistically significant differences between the two methods (*: p-value < 0.05; **: p-value < 0.01). The results shown in table I indicate that measurements of LI , LIII , and LV by the RSA and I-RSA methods are very similar, with maximum absolute differences of 0.02 ms between both estimates (on average). This 21

analysis also shows that the differences between estimates of AI , AIII , and AV by the two methods are less than 10%, except for the estimation of AI at 167 and 250 Hz, and

W

AIII at 250 Hz, in which RSA overestimates the amplitude by a factor of 15%, 14%, and

VI E

18%, respectively. The same analysis with MLR signals (table II) reveals that the latency

estimates for all parameters at all stimulation rates obtained by the RSA and I-RSA are comparable. Only the estimation of LPb at 100 Hz shows statistically significant differences

RE

of 2.44 ms, on average, between the two methods. In contrast, analysis of ANa −Pa and ANb −Pb shows significant differences between the estimates of these parameters by the two

ER

methods. At 20 Hz, there are no significant differences in the estimation of ANa −Pa , but RSA overestimates ANb −Pb by a factor of 41%, on average. At 40 Hz, RSA overestimates

ND

both ANa −Pa and ANb −Pb by a factor of 36% and 48%, respectively, as a consequence of the overlapping of the resonating Pb component on the Pa component (Bohorquez and Ozdamar,

-U

2008). At 67 Hz and greater rates, RSA produces a statistically significant underestimation of both ANa −Pa and ANb −Pb parameters of greater than 20%. The results of this analysis are

NT

IV. DISCUSSION

IA

L

presented in graphic form in figures 5 and 6.

ID E

The RSA method consists of the average of auditory responses corresponding to a burst of stimuli whose ISI varies randomly according to a predefined probability distribution along the entire stimulation sequence (randomized stimulation). This method includes a digital

NF

blanking process that considers as null values any EEG samples contaminated by the stimulus artifact (Valderrama et al., 2012). Digital blanking entails non-uniform averaging of

CO

responses along the averaging window, which can lead to significant differences in quality between different segments of the response if the amount of jitter is not sufficiently large. The average of at least 70% of the available responses assures differences of quality of less than 1.55 dB, which may be appropriate for many applications. Long-duration blanking used for long-duration stimuli (e.g., windowed tones) would impose the restriction of more jitter in 22

order to meet this requirement, which could restrict the implementation of this method in certain scenarios. Moreover, as RSA does not perform deconvolution, this method must deal

W

with interference derived from overlapping adjacent responses. This type of interference can

VI E

be reduced by averaging provided that the jitter of the stimulation sequence is larger than the dominant period of the ABR/MLR components (2 ms for ABRs and 25 ms for MLRs),

thus enabling positive and negative components of this interference to be cancelled out.

RE

When the premise of a minimum amount of jitter is fulfilled, RSA has proved to be effective

in recording ABR signals at high stimulation rates (Valderrama et al., 2012, 2014a,b,c). The

ER

implementation of the RSA method to record MLR signals is constrained by a minimum jitter of 25 ms. This long jitter has certain disadvantages: on the one hand, RSA could not

ND

be used to record MLR at rates higher than 80 Hz (using stimulation sequences ISI0−25 ), while, on the other, such high-jittered distributions may evoke auditory responses of differ-

-U

ent morphology, leading to inaccurate MLR signals, as invariance of the response over time would be wrongly assumed (Jewett et al., 2004; Ozdamar and Bohorquez, 2006; Valderrama

L

et al., 2012, 2014a).

IA

The purpose of this paper is to analyze the effects of interference associated with over-

NT

lapping responses in ABR and MLR signals recorded at different rates with randomized stimulation sequences whose jitter distributions are greater and lower than the dominant

ID E

period of the ABR/MLR components. In addition, this paper presents a new approach of the RSA method that estimates and subtracts the interference associated with overlapping responses through an iterative process in the time domain. We have called this approach

NF

“iterative-randomized stimulation and averaging” (I-RSA). I-RSA includes estimation of the interference associated with overlapping responses, subtraction of this interference from the

CO

EEG recorded, and re-estimation of the AEP. Each iteration leads to a better estimate of the AEP. The speed of convergence of this method depends on (a) correction factor α, which may solve certain problems of instability, (b) the amount of jitter of the stimulation sequence and (c) the averaged ISI compared with the length of the averaging window. Although the decision to stop iterating in I-RSA could be set automatically, depending on whether the 23

differences between two consecutive AEP estimations are negligible, in this study we found it appropriate to use 20 iterations in the implementation of I-RSA. As in RSA, the generation

W

process of stimulation sequences in I-RSA is based on randomized stimulation, where the

VI E

ISI of the stimuli varies randomly according to a predefined probability distribution (Valder-

rama et al., 2012). Two experiments were performed in this study with the dual purpose of

(a) analyzing the effects of the interference derived from overlapping responses in ABR and

RE

MLR signals obtained at different rates by the RSA and I-RSA methods, using stimulation sequences of a jitter greater and shorter than the dominant period of the ABR/MLR com-

ER

ponents; and (b) validating the I-RSA method proposed to record ABR and MLR signals at high stimulation rates.

ND

The first experiment analyzes the interference associated with overlapping responses in both real and computer simulated ABR and MLR signals obtained by the RSA and I-

-U

RSA methods at different stimulation rates, using stimulation sequences of jitter greater than (4 ms for ABRs and 50 ms for MLRs), equal to (2 ms for ABRs and 25 ms for

L

MLRs), and shorter than (0.6 ms for ABRs and 16 ms for MLRs) the dominant period of

IA

the ABR/MLR components. In the simulation framework, different EEGs were artificially

NT

synthesized by the convolution of a template waveform (an ABR and an MLR signal of high quality) with each randomized stimulation sequence, and the AEPs were obtained by the

ID E

two methods. The differences between the obtained AEPs and their corresponding template are a direct consequence of the effects of the interference due to overlapping responses, since the simulation nature of this experiment does not consider any other types of artifacts

NF

typically present in the recording of real AEPs, such as myogenic activity of the subject or electromagnetic interference. The study with simulated signals shows that the RSA method

CO

accurately estimates the templates in ABR and MLR signals at all rates when the jitter is larger than the dominant period of the ABR/MLR components, with error estimates of less than 0.02 µVRM S , indicating that the RSA method is able to reduce the effects of the interference associated with overlapping responses with averaging. However, when the jitter is shorter than the dominant period of the ABR/MLR components, this analysis 24

revealed significant differences between the template and the signals estimated by RSA as a consequence of interference associated with overlapping responses. The most relevant effects

W

of the interference due to overlapping responses in ABR signals using a jitter of 0.6 ms are

VI E

the generation of additional peaks outside the averaging window when the stimulation rate is lower than 100 Hz; and overestimation of waves I and VII at 125 Hz, waves III, VI, and VI at 167 Hz, waves I, V, and VII at 250 Hz, and wave II at 300 Hz. Additionally,

RE

the use of blanking of 1 ms duration and a jitter of 0.6 ms produced significant differences in quality between different segments of the response, and certain segments could not be

ER

obtained because of the limitation imposed by the digital blanking process. In MLR signals, the effects of this interference are overestimation of the amplitude of the Pb component at

ND

20 Hz, overestimation of all components at 40 Hz, the generation of an additional peak at about 80 ms at 20 and 40 Hz, and underestimation of the amplitude of all parameters at 67 Hz

-U

and higher rates. These results indicate that estimation of the AEP by the RSA method is not reliable when the jitter of the stimulation sequence is shorter than the dominant

L

period of the ABR/MLR components, since the interference associated with overlapping

IA

responses could not be reduced by averaging. According to the I-RSA method, this analysis shows that this method performs a highly accurate estimate of the template in both ABR

NT

and MLR signals at all rates and under all jittering conditions, which indicates that the I-RSA method is barely affected by interference associated with overlapping responses when

ID E

the jitter is greater than, equal to, or shorter than the dominant period of the ABR/MLR components. Analysis of real ABR/MLR signals is consistent with the computer-simulated

NF

study: (a) no significant differences between ABR and MLR signals obtained by the RSA and I-RSA methods are observed when the jitter of the stimulation sequence is at least equal

CO

to the dominant period of the ABR/MLR signals; and (b) when this premise is not fulfilled, the AEP obtained by RSA is contaminated by interference due to overlapping adjacent responses. The second experiment analyzes the latencies and amplitudes of the main components of real ABR and MLR signals obtained with the RSA and I-RSA methods at different 25

stimulation rates in two sets of 8 normal-hearing subjects (16 participants). The results of this analysis obtained by the I-RSA method are in accordance with previous literature. The

W

morphology of the recorded signals and the results of the analysis of latencies and amplitudes

VI E

indicate that I-RSA allows accurate AEPs to be recorded using randomized stimulation sequences with jitter distributions both greater and lower than the dominant period of

the ABR/MLR components. Furthermore, this experiment has revealed the limitations of

RE

the RSA method when the jitter of the stimulation sequences is shorter than the dominant

period of the ABR/MLR components. This study also includes an analysis of the differences

ER

between the estimation of latencies and ratio of amplitudes as carried out by the I-RSA and RSA methods. The results of this study are in accordance with the results obtained in

ND

the first experiment. These results show that (a) estimates of real ABR signals by the I-RSA and RSA methods are very similar, suggesting that RSA can be used efficiently in

-U

applications that allow the use of a jitter higher than the dominant period of the ABR/MLR components, and (b) estimates of real MLR signals by the two methods present significant

L

differences owing to the low jitter relative to the dominant period of the MLR components

NT

overlapping responses.

IA

in RSA, which does not allow the reduction by averaging of the interference associated with

The results of this study indicate that the RSA method can be efficiently used for

ID E

recording AEPs provided that the amount of jitter of the randomized stimulation sequence is greater than the dominant period of the ABR/MLR components. When this premise is not fulfilled, positive and negative components of the interference associated with overlapping

NF

responses cannot be cancelled out in the averaging process, and the resulting AEP will not be reliable. The limitation of a minimum amount of jitter in RSA is not a significant constraint

CO

in the recording of ABR signals. In ABRs, the jitter must be greater than 2 ms in order for the response to be estimated accurately by the RSA method. Theoretically, this amount of jitter would allow the generation of stimulation sequences to record ABRs at rates of up to 1000 Hz (using randomized stimulation sequences ISI0−2 ). However, the longer dominant period of the components of MLR signals entails the use of a minimum jitter of 25 ms, which 26

constrains the use of the RSA method to rates of up to 80 Hz (using randomized stimulation sequences ISI0−25 ). This stimulation rate may prove to be insufficiently high in certain

W

applications. The I-RSA approach presented here would seem to be an efficient alternative

VI E

to RSA when the amount of jitter used in the stimulation sequences is shorter than the dominant period of the ABR/MLR components. In this study, AEPs were successfully recorded by I-RSA at remarkably high stimulation rates: ABR signals were recorded at

RE

rates of up to 300 Hz and MLR signals were recorded for the first time with a method based on randomized stimulation at rates of up to 125 Hz. The performance of I-RSA maintains

ER

all the advantages of RSA: (a) it allows the jitter distribution to be controlled with precision, (b) stimulation sequences are easy to generate, and (c) it allows responses to be processed

ND

separately. Additionally, I-RSA is not constrained by the restriction of a minimum amount

NT

Acknowledgments

IA

L

-U

of jitter. These advantages may prove to be of value in various research applications.

ID E

The authors thank the subjects who participated in this study for their cooperation. This paper has been supported by the research project “Dise˜ no, implementación y evaluación de un sistema avanzado de registro de potenciales evocados auditivos del tronco (PEAT) basado

NF

en se˜ nalizacin codificada” (TEC2009-14245), R&D National Plan (2008-2011), Ministry of Finance and Competition (Government of Spain) and “European Regional Development

CO

Fund Programme” (2007-2013); by the “Granada Excellence Network of Innovation Laboratories - Startup Projects for Young Researchers Programme” (GENIL-PYR 2014), Campus of International Excellence, Ministry of Finance and Competition (Government of Spain); and by the grant “Programa de Formación de Profesorado Universitario (FPU)” (AP20093150), Ministry of Education, Culture, and Sport (Government of Spain). 27

References

W

Bardy, F., Dillon, H., and Van Dun, B. (2014a). “Least-squares deconvolution of evoked potentials and sequence optimization for multiple stimuli under low-jitter conditions,” Clinical

VI E

Neurophysiology 125, 727-737.

Bardy, F., Van Dun, B., Dillon, H., and McMahon, C. M. (2014b). “Deconvolution of over-

RE

lapping cortical auditory evoked potentials recorded using short stimulus onset-asynchrony ranges,” Clinical Neurophysiology 125, 814-826.

ER

Bell, S. L., Allen, R., and Lutman, M. E. (2001). “The feasibility of maximum length sequences to reduce acquisition time of the middle latency response,” Journal of the Acoustical

ND

Society of America 109, 1073-1081.

Bell, S. L., Allen, R., and Lutman, M. E. (2002). “Optimizing the acquisition time of

-U

the middle latency response using maximum length sequences and chirps,” Journal of the Acoustical Society of America 112, 2065-2073.

Bohorquez, J., and Ozdamar, O. (2006). “Signal to noise ratio analysis of maximum length

IA

of America 119, 2881-2888.

L

sequence deconvolution of overlapping evoked potentials,” Journal of the Acoustical Society

NT

Bohorquez, J., and Ozdamar, O. (2008). “Generation of the 40-Hz auditory steady-state response (ASSR) explained using convolution,” Clinical Neurophysiology 119, 2598-2607.

ID E

Bohorquez, J., Ozdamar, O., McNeer, R., and Morawski, K. (2009). “Clinical applications of evoked potential continuous loop averaging deconvolution (CLAD),” in 25th Southern Biomedical Engineering Conference, IFMBE Proceedings, pp. 133-134.

NF

Burkard, R., Shi, Y., and Hecox, K. E. (1990). “A comparison of maximum length and

CO

Legendre sequences for the derivation of brain-stem auditory-evoked responses at rapid rates of stimulation,” Journal of the Acoustical Society of America 87, 1656-1664.

Burkard, R. F., and Don, M. (2007). “The auditory brainstem response,” in Auditory evoked potentials: basic principles and clinical application, edited by P. Sabatini, A. M. Klinger, and J. P. Ajello (Lippincott Williams & Wilkins, Baltimore), pp. 229-253.

28

Bush, M. L., Jones, R. O., and Shinn, J. B. (2008). “Auditory brainstem response threshold differences in patients with vestibular schwannoma: a new diagnostic index,” Ear, Nose and

W

Throat Journal 87, 458-462.

VI E

de Boer, J., Brennan, S., Lineton, B., Stevens, J., and Thornton, A. R. D. (2007). “Clickevoked otoacoustic emissions (CEOAEs) recorded from neonates under 13 hours old using

conventional and maximum length sequence (MLS) stimulation,” Hearing Research 233,

RE

86-96.

Delgado, R. E., and Ozdamar, O. (1994). “Automated auditory brainstem response inter-

ER

pretation,” IEEE Engineering in Medicine and Biology Magazine 13, 227-237.

Delgado, R. E., and Ozdamar, O. (2004). “Deconvolution of evoked responses obtained at

ND

high stimulus rates,” Journal of the Acoustical Society of America 115, 1242-1251. Don, M., Allen, A., and Starr, A. (1977). “Effect of click rate on the latency of auditory

-U

brain stem responses in humans,” Annals of Otology, Rhinology and Laryngology 86, 186195.

L

Eggermont, J. J. (2007). “Electric and Magnetic Fields of Synchronous Neural Activity:

IA

Peripheral and Central Origins of AEPs,” in Auditory evoked potentials: basic principles and clinical application, edited by P. Sabatini, A. M. Klinger, and J. P. Ajello (Lippincott

NT

Williams & Wilkins, Baltimore), pp. 2-21. Elberling, C., and Don, M. (2007). “Detecting and assessing synchronous neural activity

ID E

in the temporal domain (SNR, Response detection),” in Auditory evoked potentials: basic principles and clinical application, edited by P. Sabatini, A. M. Klinger, and J. P. Ajello

NF

(Lippincott Williams & Wilkins, Baltimore), pp. 102-123. Eysholdt, U., and Schreiner, C. (1982). “Maximum length sequences: A fast method for

CO

measuring brain-stem-evoked responses,” Audiology 21, 242-250. Fifer, R. C., and Sierra-Irizarry, B. (1988). “Clinical applications of the auditory middle

latency response,” The American Journal of Otology 9, 47-56. Galambos, R., Makeig, S., and Talmachoff, P. J. (1981). “A 40-Hz auditory potential recorded from the human scalp,” in Proceedings of the National Academy of Sciences of 29

the United States of America 78, pp. 2643-2647.

40-Hz steady-state fields evoked by two-tone sequences,” Hearing Research 257, 83-92.

W

Gutschalk, A., Oldermann, K., and Rupp, A. (2009). “Rate perception and the auditory

VI E

Hall, J. W. (2007). New Handbook of Auditory Evoked Responses, 1st ed. (Allyn and Bacon, Boston, MA), p. 750.

Hine, J. E., Ho, C.-T., Slaven, A., and Thornton, A. R. D. (2001). “Comparison of transient

RE

evoked otoacoustic emission thresholds recorded conventionally and using maximum length sequences,” Hearing Research 156, 104-114.

ER

Jewett, D. L., and Williston, J. S. (1971). “Auditory-evoked far fields averaged from the scalp of humans,” Brain 94, 681-696.

ND

Jewett, D. L., Caplovitz, G., Baird, B., Trumpis, M., Olson, M. P., and Larson-Prior, L. J. (2004). “The use of QSD (q-sequence deconvolution) to recover superposed, transient

-U

evoked-responses,” Clinical Neurophysiology 115, 2754-2775.

Jiang, Z. D., Brosi, D. M., Shao, X. M., and Wilkinson, A. R. (2000). “Maximum Length

L

Sequence brainstem auditory evoked responses in term neonates who have perinatal hypoxia-

IA

ischemia,” Pediatric Research 48, 639-645.

Jiang, Z. D., Wu, Y. Y., and Wilkinson, A. R. (2009). “Age-related changes in BAER at

NT

different click rates from neonates to adults,” Acta Paediatrica, International Journal of Paediatrics 98, 1284-1287.

ID E

Kacker, S. K., and Deka, R. C. (1986). “Auditory brainstem evoked responses in Meniere’s disease,” Indian Journal of Otolaryngology 38, 14-17.

NF

Kjaer, M. (1980). “Brain stem auditory and visual evoked potentials in multiple sclerosis,” Acta Neurologica Scandinavica 62, 14-19.

CO

Lasky, R. E. (1984). “A developmental study on the effect of stimulus rate on the auditory evoked brain-stem response,” Electroencephalography and Clinical Neurophysiology 59, 411419. Lasky, R. E. (1997). “Rate and adaptation effects on the auditory evoked brainstem response in human newborns and adults,” Hearing Research 111, 165-176. 30

Lavoie, B. A., Barks, A., and Thornton, A. R. D. (2010). “Linear and nonlinear temporal interaction components of mid-latency auditory evoked potentials obtained with maximum

W

length sequence stimulation,” Experimental Brain Research 202, 231-237.

VI E

Leung, S., Slaven, A., Thornton, A. R. D., and Brickley, G. J. (1998). “The use of high

stimulus rate auditory brainstem responses in the estimation of hearing threshold,” Hearing Research 123, 201-205.

RE

Lina-Granade, G., Collet, L., Morgon, A., and Salle B. (1993). “Maturation and effect of

stimulus rate on brainstem auditory evoked potentials,” Brain and Development 15, 263-269.

(Prentice-Hall, Upper Saddle River, NJ), p. 870.

ER

Oppenheim, A.V., and Schafer, R.W. (1999). Discrete-Time Signal Processing, 2nd ed.

ND

Ozdamar, O., Delgado, R. E., Yavuz, E., Thombre, K., and Anderson, M. 2003a. “Acquisition of ABRs at very high stimulation rate using CLAD (Continuous Loop Algorithm

-U

Deconvolution),” ARO Abstract 40, 44.

Ozdamar, O., Delgado, R. E., Yavuz, E., Thombre, K., and Acikgoz, N. 2003b. “Deconvo-

L

lution of auditory evoked potentials obtained at high stimulus rates,” in Proceedings of the

IA

1st International IEEE EMBS Conference on Neural Engineering, pp. 285-288. Ozdamar, O., and Bohorquez, J. (2006). “Signal-to-noise ratio and frequency analysis of

NT

continuous loop averaging deconvolution (CLAD) of overlapping evoked potentials,” Journal of the Acoustical Society of America 119, 429-438.

ID E

Ozdamar, O., Bohorquez, J., and Ray, S. S. (2007). “Pb (P1 ) resonance at 40 Hz: Effects of high stimulus rate on auditory middle latency responses (MLRs) explored using deconvolu-

NF

tion,” Clinical Neurophysiology 118, 1261-1273. Pastor, M. A., Artieda, J., Arbizu, J., Marti-Climent, J. M., Penuelas, I., and Masdeu, J.

CO

C. (2002). “Activation of human cerebral and cerebellar cortex by auditory stimulation at 40 Hz,” Journal of Neuroscience 22, 10501-10506. Picton, T. W., Champagne, S. C., and Kellet, A. J. C. (1992). “Human auditory evoked potentials recorded using maximum length sequences,” Electroencephalography and Clinical Neurophysiology 84, 90-100. 31

Podoshin, L., Ben-David, Y., Pratt, H., Fradis, M., and Feiglin, H. (1986). “Noninvasive recordings of cochlear evoked potentials in Meniere’s disease,” Archives of Otolaryngology -

W

Head and Neck Surgery 112, 827-829.

VI E

Pratt, H., and Sohmer, H. (1976). “Intensity and rate functions of cochlear and brainstem evoked responses to click stimuli in man,” Archives of Oto-Rhino-Laryngology 212, 85-92.

Pratt, H. (2007). “Middle-Latency Responses,” in Auditory evoked potentials: basic princi-

RE

ples and clinical application, edited by P. Sabatini, A. M. Klinger, and J. P. Ajello (Lippincott Williams & Wilkins, Baltimore), pp. 463-481.

ER

Pratt, H. (2011). “Sensory ERP components,” in Oxford Handbook of Event-Related Potential Components, edited by E. S. Kappenman and S. J. Luck (Oxford University Press, New

ND

York), pp. 1-35.

Presacco, A., Bohorquez, J., Yavuz, E., and Ozdamar, O. (2010). “Auditory steady-state

-U

responses to 40 Hz click trains: Relationship to middle latency, gamma band and beta band responses studied with deconvolution,” Clinical Neurophysiology 121, 1540-1550.

L

Rudell, A. (1987). “A fiber tract model of auditory brain-stem responses,” Electroen-

IA

cephalography and Clinical Neurophysiology 67, 53-62. Stockard, J. J., Stockard, J. E., and Sharbrough, F. W. (1978). “Nonpathologic factors

18, 177-209.

NT

influencing brainstem auditory evoked potentials,” American Journal of EEG Technology

ID E

Stone, J. L., Calderon-Arnulphi, M., Watson, K. S., Patel, K., Mander, N. S., Suss, N., Fino, J., and Hughes, J. R. (2009). “Brainstem Auditory Evoked Potentials-A review and

NF

modified Studies in Healthy Subjects,” Journal of Clinical Neurophysiology 26, 167-175. Thornton, A. R. D., and Slaven, A. (1993), “Auditory brainstem responses recorded at

CO

fast stimulation rates using maximum length sequences”, British Journal of Audiology 27, 205-210. Thornton, A. R. D., Lineton, B., Baker, V. J., and Slaven, A. (2006). “Nonlinear properties of otoacoustic emissions in normal and impaired hearing,” Hearing Research 219, 56-65. Thornton, A. R. D. (2007). “Instrumentation and recording parameters,” in Auditory evoked 32

potentials: basic principles and clinical application, edited by P. Sabatini, A. M. Klinger, and J. P. Ajello (Lippincott Williams & Wilkins, Baltimore), pp. 73-101.

W

Tuliani, J. (2001). “De Bruijn sequences with efficient decoding algorithms,” Discrete Math-

VI E

ematics 226, 313-336.

Valderrama, J. T., Alvarez, I., de la Torre A., Segura, J. C., Sainz, M., and Vargas, J. L.

Technics Technologies Education Management 6, 876-889.

RE

(2011). “Educational approach of a BAER recording system based on experiential learning,”

Valderrama, J. T., Alvarez, I., de la Torre A., Segura, J. C., Sainz, M., and Vargas, J.

ER

L. (2012). “Recording of auditory brainstem response at high stimulation rates using randomized stimulation and averaging,” Journal of the Acoustical Society of America 132,

ND

3856-3865.

Valderrama, J. T., de la Torre A., Alvarez, I., Segura, J. C., Thornton, A. R. D., Sainz,

-U

M., and Vargas, J. L. (2013). “Auditory middle latency responses recorded at high stimulation rates using randomized stimulation and averaging,” in International Evoked Response

L

Audiometry Study Group (IERASG) meeting, New Orleans, LA (June 9-13, 2013), pp. 56.

IA

Valderrama, J. T., de la Torre A., Alvarez, I., Segura, J. C., Thornton, A. R. D., Sainz, M., and Vargas, J. L. (2014a). “A study of adaptation mechanisms based on ABR recorded at

NT

high stimulation rate,” Clinical Neurophysiology 125, 805-813. Valderrama, J. T., de la Torre A., Alvarez, I., Segura, J. C., Thornton, A. R. D., Sainz, M.,

ID E

and Vargas, J. L. (2014b). “Automatic quality assessment and peak identification of auditory brainstem responses with fitted parametric peaks,” Computer Methods and Programs

NF

in Biomedicine 114, 262-275. Valderrama, J. T., de la Torre A., Alvarez, I., Segura, J. C., Sainz, M., and Vargas, J. L.

CO

(2014c). “A flexible and inexpensive high-performance auditory evoked response recording system appropriate for research purposes,” Biomedizinische Technik/Biomedical Engineering, DOI: http://dx.doi.org/10.1515/bmt-2014-0034. Wang, T., Huang, J.-H., Lin, L., and Zhan, C. A. (2013). “Continuous- and discretetime stimulus sequences for high stimulus rate paradigm in evoked potential stud33

ies,” Computational and Mathematical Methods in Medicine, article ID 396034, DOI: http://dx.doi.org/10.1155/2013/396034.

VI E

use of noise estimate,” Electroencephalography and Clinical Neurophysiology 50, 25-34.

W

Wong, P. K. H., and Bickford R. G. (1980). “Brain stem auditory evoked potentials: the

Yagi, T., and Kaga, K. (1979). “The effect of the click repetition rate on the latency of the auditory evoked brain stem response and its clinical use for a neurological diagnosis,”

RE

Archives of Oto-Rhino-Laryngology 222, 91-96.

Zollner, C., Karnahl, T., and Stange, G. (1976). “Input-output function and adaptation

ER

behaviors of the five early potentials registered with the earlobe-vertex pick up,” Archives

CO

NF

ID E

NT

IA

L

-U

ND

of Oto-Rhino-Laryngology 212, 23-33.

34

Stimulation rate

LI

LIII

LV

AI

AIII

AV

-0.01 (0.02) 0.00 (0.00) 0.00 (0.01) 0.05 (0.04)* 0.00 (0.01) -0.02 (0.02)**

56 Hz

0.00 (0.00) 0.00 (0.01) -0.01 (0.00)* 0.05 (0.05)* 0.00 (0.01) -0.02 (0.01)**

71 Hz

0.00 (0.01) 0.00 (0.01) 0.00 (0.01)

0.05 (0.04) -0.01 (0.01) -0.02 (0.01)**

83 Hz

0.00 (0.01) 0.00 (0.01) 0.00 (0.00)

0.05 (0.05)

VI E

0.00 (0.01) -0.02 (0.01)**

RE

100 Hz

W

45 Hz

-0.01 (0.01)* 0.00 (0.01) 0.00 (0.00) 0.07 (0.05)* -0.01 (0.02)* -0.04 (0.03)** 0.00 (0.01) -0.01 (0.02) 0.00 (0.01)

0.02 (0.06)

0.03 (0.06)

-0.02 (0.07)

167 Hz

0.00 (0.01) 0.00 (0.01) 0.01 (0.03)

0.15 (0.14) -0.03 (0.13)

0.06 (0.07)

250 Hz

0.00 (0.01) 0.00 (0.01) 0.02 (0.05)

ND

ER

125 Hz

0.14 (0.10)

0.18 (0.17)

0.07 (0.05)*

-U

TABLE I. Mean (and standard deviation in parentheses) of the differences in latencies expressed in milliseconds (LRSA −LI−RSA ) and the ratio of amplitudes (estimated as ARSA /AI−RSA − 1) between ABR signals obtained by the RSA and I-RSA methods. Sta-

L

tistically significant differences between the two methods are expressed with asterisks (*: p-

CO

NF

ID E

NT

IA

value < 0.05; **: p-value < 0.01).

35

Stimulation rate

LNa

LPa

LNb

LPb

ANa −Pa

ANb −Pb

0.00 (0.06) -0.04 (0.07) 0.45 (1.31) -0.05 (0.14)

0.00 (0.00)

0.00 (0.00)

20 Hz

-0.08 (0.21) -0.03 (0.10) 0.11 (0.22) -0.19 (0.26)

0.05 (0.05)

0.41 (0.26)*

40 Hz

-0.24 (0.64) 0.09 (0.22) 0.65 (1.44) 0.06 (0.44) 0.36 (0.10)** 0.48 (0.21)**

67 Hz

0.16 (0.31) -0.30 (0.43) -0.30 (0.76) -1.31 (1.59) -0.21 (0.06)** -0.27 (0.10)**

100 Hz

0.13 (0.24) 0.07 (0.42) -0.76 (1.18) -2.44 (2.72)* -0.27 (0.06)** -0.33 (0.05)**

125 Hz

-0.03 (0.22) -0.05 (0.50) -0.37 (0.92) -0.41 (1.22) -0.23 (0.11)** -0.27 (0.08)**

ER

RE

VI E

W

8 Hz

TABLE II. Mean (and standard deviation in parentheses) of the differences in laten-

ND

cies expressed in milliseconds (LRSA −LI−RSA ) and the ratio of amplitudes (estimated as ARSA /AI−RSA − 1) between MLR signals obtained by the RSA and I-RSA methods. Sta-

CO

NF

ID E

NT

IA

L

value < 0.05; **: p-value < 0.01).

-U

tistically significant differences between the two methods are expressed with asterisks (*: p-

36

Influence of the amount of jitter of the stimulation sequence and the duration

RE

FIG. 1

VI E

W

List of Figures

of digital blanking on the number of samples averaged along the averaging window for an ABR signal with the RSA method. This figure shows that

ER

low-jittered sequences and long-duration blanking can produce appreciable 39

-U

ND

differences in terms of quality between different segments of the response. . .

Autocorrelation function for given high-quality ABR and MLR signals. This

L

FIG. 2

IA

figure shows that the dominant period of ABR and MLR signals is about 40

ID E

NT

2 ms and 25 ms, respectively. . . . . . . . . . . . . . . . . . . . . . . . . . .

An illustration of estimation of the AEP based on I-RSA. Sampling frequency:

NF

FIG. 3

CO

fS = 25 kHz; length of the averaging window: J = 2500 samples (100 ms); correction factor: α = 1. (A) Transient AEP. This example shows a MLR

signal. (B) Synchronization signal and raw EEG (in this example, the EEG was synthesized from a real MLR response). (C) Estimation of the AEP on the preceding iteration. (D) EEG with all overlapping responses subtracted. (E) Estimation of the AEP on the iteration i. . . . . . . . . . . . . . . . . . 37

41

FIG. 4

(color online) Real and computer-simulated ABR and MLR signals obtained

quences of jitter distributions that are greater than, equal to, and shorter

VI E

than the dominant period of the ABR/MLR components. In the study with

W

with RSA and I-RSA at different stimulation rates using stimulation se-

simulated ABR/MLR signals, the interference associated with overlapping

responses has been estimated as the RMS value of the difference between the

RE

template and the signals obtained. This study indicates that (a) the I-RSA method estimates the response accurately at all rates and under all jittering

ER

conditions; (b) the RSA method requires an amount of jitter at least equal to the dominant period of the recorded AEP to allow the interference associ-

ND

ated with overlapping responses to be suppressed by averaging; and (c) the digital blanking process in RSA may constrain the use of short distributions

-U

of the jitter, e.g., the ABR signals corresponding to a jitter of 0.6 ms present appreciable differences in quality along the averaging window, and certain

(color online) Upper panel: ABR signals obtained in a set of 8 normal-hearing

IA

FIG. 5

42

L

segments of the response could not be obtained. . . . . . . . . . . . . . . . .

subjects at different stimulation rates obtained by RSA and I-RSA using

NT

stimulation sequences of a jitter of 4 ms (greater than the dominant period of ABR components). Lower panel: MLR signals obtained in a different set

ID E

of 8 normal-hearing subjects at different stimulation rates obtained by RSA and I-RSA using stimulation sequences of a jitter of 16 ms (lower than the 43

NF

dominant period of MLR components). . . . . . . . . . . . . . . . . . . . . .

CO

FIG. 6

(color online) Mean (and standard deviation in errorbars) of the latencies

and amplitudes of the main components of ABR and MLR signals obtained by the RSA and I-RSA methods at different stimulation rates. . . . . . . . .

38

44

ISI5−7 stimulation sequence 20

20

18

18

18

16

16

70%

14

16

70%

14

12

12

12

10

10

10

8

8

8

6

6

6

4

4

4

2

2

Blanking = 1 ms

0 0

1

2

3

4

5

6

7

8

9

10

2

Blanking = 1 ms

0 0

1

2

70%

14

3

4

Blanking = 1 ms

0 5

6

7

8

9

10

0

1

Averaging window (ms)


2

16

16

70%

12

10

10

8

8

6

6

4

4

2

2

Blanking = 2 ms

0 0

1

2

3

4

5

7

8

9

10

10 8 6 4 2

Blanking = 2 ms

0

1

2

3

4

5

6

7

9

10

IA NT

10 8 6 4 2

Blanking = 0.5 ms

0

0

1

2

3

4

5

6

7

8

10

3

4

5

6

7

8

9

10

18 16 70%

14

12

12

10

10

8

8

6

6

4

4

2

2 Blanking = 0.5 ms

0 0

1

2

3

4

5

Blanking = 0.5 ms

0 6

7

8


9

10

0

1

2

3

4

5

6

7

8

9

10


NF


9

2

20

70%

14

12

1

ISI5−5.5 stimulation sequence

L

ISI5−7 stimulation sequence

16

14

0


18

70%

Blanking = 2 ms

0 8

20

16

70%

14


18

10

12


ID E

Number of averaged samples (x1000)

16

70%


20

9

18

0 6

8

20

14

12

7

ND

14

6

ER

18

-U


20

18

5



20

4


Influence of the duration of blanking ISI5−9 stimulation sequence

3

VI E

20

W


RE


Influence of the amount of jitter ISI5−9 stimulation sequence

FIG. 1. Influence of the amount of jitter of the stimulation sequence and the duration

CO

of digital blanking on the number of samples averaged along the averaging window for an ABR signal with the RSA method. This figure shows that low-jittered sequences and longduration blanking can produce appreciable differences in terms of quality between different segments of the response.

39

100

100

0

−50

−100 0

50

0

−50

2

4

6

8

−100 0

10

20

40

60

Shift (ms)

80

100

RE

Shift (ms)

VI E

50

W

MLR autocorrelation 150

Percentage (%)

Percentage (%)

ABR autocorrelation 150

ER

FIG. 2. Autocorrelation function for given high-quality ABR and MLR signals. This figure shows that the dominant period of ABR and MLR signals is about 2 ms and 25 ms,

CO

NF

ID E

NT

IA

L

-U

ND

respectively.

40

(A) x(j) − Transient AEP µV

0.5 0 −0.5 0

1250

J = 2500

W

Sample (j)

m1 = 84

1 0.5 0 −0.5 0

m2 = 1416

1000

m3 = 2793

2000

3000

m4 = 4060

4000

m5

5000

VI E

µV

(B) y(n) − Electroencephalogram, s(n) − Synchronization signal

6000

Sample (n)

RE

(C) hi−1(j) − AEP estimation, preceding iteration µV

0.5 0 −0.5 0

1250

J = 2500

µV

m1

1 0.5 0 −0.5

1000

m3

2000

m4

3000

m5

4000

5000

6000

ND

0

m2

ER

Sample (j)

(D) σ(n) = y(n) − s(n) ∗ hi−1(j)

Sample (n)

(E) hi(j) − AEP estimation, iteration i

-U

hi−1(j)

+

σ2 = σ(j+m(2))

σ3 = σ(j+m(3))

L

0.5 µV

...

NT

IA

...

1250

σK = σ(j+m(K))

hi(j) J = 2500

Sample (j)

ID E

0

Average

σ1 = σ(j+m(1))

FIG. 3. An illustration of estimation of the AEP based on I-RSA. Sampling frequency:

NF

fS = 25 kHz; length of the averaging window: J = 2500 samples (100 ms); correction factor: α = 1. (A) Transient AEP. This example shows a MLR signal. (B) Synchronization

CO

signal and raw EEG (in this example, the EEG was synthesized from a real MLR response). (C) Estimation of the AEP on the preceding iteration. (D) EEG with all overlapping responses subtracted. (E) Estimation of the AEP on the iteration i.

41

ABR

167 Hz 125 Hz 100 Hz 83 Hz 71 Hz V I III

0.3 µV

Template 10

15

20

250 Hz 167 Hz 125 Hz 100 Hz 83 Hz 71 Hz V I III

25

0

5

10

15

125 Hz

RSA: 0.0616 I−RSA: 0.0018

100 Hz

RSA: 0.0498 I−RSA: 0.0019

Time (ms)

Jitter 4 ms

Jitter 2 ms

25

0

250 Hz

167 Hz

167 Hz

125 Hz

125 Hz

100 Hz

100 Hz

15

20

25

125 Hz 100 Hz

83 Hz

0.3 µV

71 Hz

25

20

ER

250 Hz

167 Hz

71 Hz

15

Real ABRs

250 Hz

0.3 µV

10

Jitter 0.6 ms

300 Hz

10

5

Real ABRs

83 Hz

0.3 µV

Template

300 Hz

5

RSA: 0.0441 I−RSA: 0.0025

71 Hz

300 Hz

0

RSA: 0.0445 I−RSA: 0.0019

83 Hz

Time (ms)

Time (ms)

Real ABRs

Stimulation rate

20

167 Hz

RSA: 0.0629 I−RSA: 4·10−4 RSA: 0.1190 I−RSA: 6·10−4 RSA: 0.0786 I−RSA: 0.0012

250 Hz

V I III

0.3 µV

Template

Error (µVRMS)

300 Hz

0

5

Time (ms)

10

15

83 Hz

20

0.3 µV

71 Hz

ND

Stimulation rate

250 Hz

Simulated ABRs

Error (µVRMS) RSA: 0.0171 I−RSA: 4·10−4 RSA: 0.0309 I−RSA: 3·10−4 RSA: 0.0200 I−RSA: 2·10−4 RSA: 0.0196 I−RSA: 5·10−4 RSA: 0.0158 I−RSA: 4·10−4 RSA: 0.0158 I−RSA: 5·10−4 RSA: 0.0153 I−RSA: 7·10−4

300 Hz

25

0

5

Time (ms)

10

15

20

25

Time (ms)

-U

MLR Jitter 50 ms

I−RSA

Jitter 16 ms Simulated MLRs

Simulated MLRs

RSA: 0.0771 I−RSA: 0.0098

100 Hz

RSA: 0.0869 I−RSA: 1·10−4

67 Hz

RSA: 0.0227 I−RSA: 1·10−4

67 Hz

RSA: 0.0788 I−RSA: 6·10−4

40 Hz

RSA: 0.0149 I−RSA: 0.0025

40 Hz

RSA: 0.0219 I−RSA: 5·10−4

40 Hz

RSA: 0.1154 I−RSA: 0.0028

20 Hz

RSA: 0.0186 I−RSA: 1·10−4

L

Error (µVRMS)

Error (µVRMS)

125 Hz

Error (µVRMS)

20 Hz

RSA: 0.0296 I−RSA: 4·10−4

20 Hz

RSA: 0.0683 I−RSA: 0.0021

8 Hz

RSA: 4·10−15 I−RSA: 4·10−14

IA

Stimulation rate

Simulated MLRs

RSA: 4·10−15 I−RSA: 5·10−14

V

Pa

1 µV

Pb

Template 0

20

Nb

40

60

80

RSA: 4·10−15 I−RSA: 5·10−14

8 Hz

V

Pa

1 µV

Pb

Template

NT

Na

100

0

Na

20

V

40

Na

60

80

100

0

20

40

60

80

Time (ms)

Time (ms)

Jitter 50 ms

Jitter 25 ms

Jitter 16 ms

ID E

125 Hz 100 Hz 67 Hz

67 Hz 40 Hz

20 Hz

20 Hz

20 Hz

1 µV

8 Hz

40

60

Time (ms)

80

100

1 µV

8 Hz 0

20

40

60

Time (ms)

80

100

1 µV

8 Hz 0

20

40

60

80

100

Time (ms)

CO

20

100

Real MLRs

Real MLRs

40 Hz

NF

1 µV

Pb Nb

Time (ms)

40 Hz

0

Pa

Template

Nb

Real MLRs

Stimulation rate

RSA

Jitter 25 ms

8 Hz

W

RSA: 0.0088 I−RSA: 9·10−4 RSA: 0.0069 I−RSA: 1·10−4 RSA: 0.0086 I−RSA: 1·10−4 RSA: 0.0079 I−RSA: 8·10−5 RSA: 0.0078 I−RSA: 7·10−5 RSA: 0.0072 I−RSA: 6·10−5 RSA: 0.0069 I−RSA: 1·10−4

5

Simulated ABRs

Error (µVRMS)

I−RSA

VI E

Simulated ABRs 300 Hz

0

RSA

Jitter 0.6 ms

Jitter 2 ms

RE

Jitter 4 ms

FIG. 4. (color online) Real and computer-simulated ABR and MLR signals obtained with RSA and I-RSA at different stimulation rates using stimulation sequences of jitter distributions that are greater than, equal to, and shorter than the dominant period of the ABR/MLR components. In the study with simulated ABR/MLR signals, the interference 42 estimated as the RMS value of the difference associated with overlapping responses has been between the template and the signals obtained. This study indicates that (a) the I-RSA method estimates the response accurately at all rates and under all jittering conditions;

ABR Subject 1

Subject 2

Subject 3

Subject 4 RSA I−RSA

VI E

W

250 Hz 167 Hz 125 Hz 100 Hz 83 Hz 71 Hz 56 Hz 45 Hz

0.4 µV 2

4

6

8 10

0

Subject 5

2

4

6

8 10

0

2

Subject 6

4

6

8 10

0

Subject 7

4

6

8 10

Subject 8

0

2

4

6

8 10

0

Time (ms)

2

4

6

8 10

0

Time (ms)

ND

ER

250 Hz 167 Hz 125 Hz 100 Hz 83 Hz 71 Hz 56 Hz 45 Hz

2

RE

0

2

4

6

8 10

Time (ms)

-U

MLR

Subject 9

Subject 10

125 Hz

Subject 11

0

2

4

6

0.4 µV 8 10

Time (ms)

Subject 12 RSA I−RSA

L

100 Hz 67 Hz

IA

40 Hz 20 Hz

NT

8 Hz

1 µV

0 20 40 60 80 100

0 20 40 60 80 100

0 20 40 60 80 100

Subject 13

Subject 14

Subject 15

Subject 16

ID E

0 20 40 60 80 100

125 Hz 100 Hz

67 Hz

NF

40 Hz 20 Hz

1 µV

CO

8 Hz 0 20 40 60 80 100

0 20 40 60 80 100

0 20 40 60 80 100

0 20 40 60 80 100

Time (ms)

Time (ms)

Time (ms)

Time (ms)

FIG. 5. (color online) Upper panel: ABR signals obtained in a set of 8 normal-hearing subjects at different stimulation rates obtained by RSA and I-RSA using stimulation sequences of a jitter of 4 ms (greater than the dominant period of ABR components). Lower panel: MLR signals obtained in a different set of 8 43 normal-hearing subjects at different stimulation rates obtained by RSA and I-RSA using stimulation sequences of a jitter of 16 ms (lower than the dominant period of MLR components).

ABR Wave V

7 6 5

Amplitudes wave III (µV)

Amplitudes wave V (µV)

0.4

0.4

0.4

0.3

0.3

0.3

0.2

0.2

0.2

0.1

0.1

0.1

RSA I−RSA

W

Amplitudes wave I (µV)

Latencies (ms)

Wave III

4

2

Wave I

1 30

90

150

210

0 30

270

90

150

210

270

0 30



90

150

210

0 30

270


Nb

30

Pa

20

Na

10 0

25

50

75

100

125

1.6

1.2

1.2

0.8

0.8

0.4 0 0

0.4

25

50

75

100


-U


RSA I−RSA

ER

40

1.6

0 0

ND

Pb

270

Amplitudes Nb−Pb (µV)

60 50

210

RE

Amplitudes Na−Pa (µV)

150


MLR Latencies (ms)

90

VI E

3

125

25

50

75

100

125


FIG. 6. (color online) Mean (and standard deviation in errorbars) of the latencies and

L

amplitudes of the main components of ABR and MLR signals obtained by the RSA and

CO

NF

ID E

NT

IA

I-RSA methods at different stimulation rates.

44

NHS 2012 Conference Villa Erba Congress Center, Cernobbio (Como Lake), Italy - June 5-7, 2012 Abstract Form Type of preferred presentation

Oral

X

Poster

Please mark your preference with X

Corresponding author  Family Name / Name Institution/Hospital/Department Address Postal code/City/Country

Valderrama Valenzuela, Joaquin T. Department of signal theory, networking and communications. CITIC-UGR. University of Granada. Spain. C\ Periodista Daniel Saucedo Aranda s/n, 18004, Granada, Spain.

phone / fax e-mail

Phone: +34 958 241 521, Fax: +34 958 240831. [email protected]

All messages will be sent ONLY to the corresponding author, please check very carefully your e-mail address Title of the abstract

Reducing recording time of brainstem auditory evoked responses by the use of randomized stimulation

Authors

Valderrama J1, Alvarez I1, de la Torre A1, Segura JC1, Sainz M2,3, and Vargas J2

1) Affiliation of all Authors Department of Signal Theory, Networking and Communications. CITIC(Institution/Hospital/Department, UGR. University of Granada. Spain. 2) City, Country) ENT Service. San Cecilio University Hospital. Granada. Spain. 3) Department of Surgery and Its Specialties. University of Granada. Spain.

Before preparing your Abstract please read carefully the instructions given on the website

Submission deadline: February 16, 2012 Text only – no figures, no formulas. Maximum 300 words TEXT: Many protocols for newborn and infant hearing screening incorporate the acquisition of Brainstem Auditory Evoked Responses (BAER). This evoked potential represents the neural activity associated with the auditory brainstem function in response to a sound stimulus. The biological response waveform is characterized by a series of positive waves that occur during the first 10msec after stimulus presentation. Since wave V is the most robust wave, it is usually used as indicator of hearing loss. The conventional acquisition technique elicit the biological response by presenting to the patient a pulse train with a fixed inter-stimulus interval (ISI) and averaging then the response to each stimulus. Usually, more than 1500 responses must be averaged to obtain a reliable biological response. The exploration time is therefore the main limitation in the recording of these potentials. This preliminary study presents a novel technique to reduce the recording time. Instead of using a pulse train with a fixed ISI, we propose to elicit the biological response by using a pulse train with a random ISI. To analyze the feasibility of the proposed technique, the BAER of four normal hearing adults were acquired considering the following inter-pulse intervals: ISI25 (conventional technique), ISI3-8 (uniformly distributed random inter-pulse interval in the range from 3 to 8msec), ISI5-10, ISI8-13, ISI10-15, ISI15-20 and ISI20-25. The recordings confirm that the hearing threshold can be determined in a shorter time using the proposed technique since the sound intensity threshold at which wave V appears is similar in both conventional and proposed techniques. Although the waves recorded with the proposed technique present lesser amplitudes and greater latencies due to adaptation, a fast detection of the wave V facilitates the implementation of newborn and infant hearing screening protocols.



REDUCING RECORDING TIME OF BRAINSTEM AUDITORY EVOKED RESPONSES BY THE USE OF RANDOMIZED STIMULATION Joaquín T. Valderrama, Isaac Álvarez, Ángel de la Torre, José Carlos Segura, Manuel Sáinz , José Luis Vargas

Department of Signal Theory, Networking and Communications. CITIC-UGR. University of Granada (Spain). ENT Service. San Cecilio University Hospital. Granada (Spain). Department of Surgery and its Specialties. University of Granada (Spain).

1 - INTRODUCTION

3 - ASSESSMENT

Many protocols for newborn and infant hearing screening incorporate the acquisition of Brainstem Auditory Evoked Responses (BAER). This evoked potential represents the neural activity associated with the brainstem in response to a sound stimulus. The biological response waveform is characterized by a series of positive waves that occur during the first 10 ms after stimulus presentation. Since wave V is the most robust wave, it is usually used as an indicator of hearing loss. The conventional acquisition technique elicit the biological response by presenting to the patient a pulse train with a fixed inter-stimulus interval (ISI) and averaging then the sweeps corresponding to each stimulus. Usually, more than 1500 sweeps must be averaged to obtain a reliable biological response. The exploration time is therefore an important limitation in the recording of these potentials. This preliminary study presents a novel technique to reduce the recording time.

2 - RANDOMIZED STIMULATION

In contrast to the conventional technique, in which stimuli are presented synchronously with a period greater than the averaging window (usually 10 ms) (figure 1.A), the Randomized Stimulation technique consists of the average of auditory responses corresponding to a burst of stimuli whose period varies randomly within two values according to a predefined probability distribution. Figure 1.B shows an example of Randomized Stimulation sequence in which the ISI varies uniformly random between 3 and 8 ms. In this example, ISI is smaller than the averaging window, leading to overlapping responses. A histogram for an ISI3-8 sequence is presented in figure 1.C.

Fig. 2. BAER signals from 4 normally hearing subjects recorded by averaging 4.000 sweeps at different stimulation rates using the Randomized Stimulation technique. Wave V can be identified in each recorded signal.

The BAER from four normally hearing subjects were acquired using the Randomized Stimulation technique at different stimulation rates in order to test the feasibility of the proposed methodology (figure 2). These recordings were obtained by averaging 4.000 sweeps. Figure 2 shows that wave V can be recognized in the four subjects at every stimulation rate. This figure also shows that the effects of adaptation on the evoked potentials increase with stimulation rate (amplitudes decrease and latencies increase, especially in more central waves), which confirms the recording of BAER. The time necessary to obtain BAER at different stimulation rates is analyzed in table 1. The results of this study show that the use of the Randomized Stimulation technique at high stimulation rates yields to an important reduction of the recording time.

ISI

Sweeps

Recording time

Time reduction

ISI25

4046

101,15 s

-

ISI20-25

4174

93,92 s

7,14 %

ISI15-20

4188

73,29 s

27,54 %

ISI10-15

4152

51,90 s

48,69 %

ISI8-13

3984

41,83 s

58,64 %

ISI3-8

3992

21,96 s

78,29 %

Table 1. Analysis of the time required to obtain BAER by averaging 4.000 sweeps at different stimulation rates using the Randomized Stimulation technique. The recording time is compared with the conventional technique at ISI = 25 ms.

4 - CONCLUSIONS ✓ The Randomized Stimulation technique can be used to record BAER at high stimulation rates.

Figure 1. (A) Stimulation sequence in the conventional technique. Stimuli are presented synchronously with a period greater than the averaging window. (B) ISI3-8 Randomized Stimulation sequence. Inter-stimulus interval (ISI) is smaller than the averaging window, leading to overlapping responses. (C) Histogram for an ISI3-8 Randomized Stimulation sequence. ISI varies uniformly random within the range 3 - 8 ms.

E-mail us to [email protected]

✓ The recording time required to obtain brainstem auditory evoked potentials can be considerably reduced using the Randomized Stimulation technique. ✓ A fast detection of the wave V facilitates the implementation of newborn and infant hearing screening protocols. June 5-7, NHS Congress 2012, Cernobbio (Lake Como), Italy

AHS 2012 Conference

Villa Erba Congress Center, Cernobbio (Como Lake), Italy - June 7-9, 2012 Abstract Form Type of preferred presentation

Corresponding author  Family Name / Name Department Address Postal code/City/Country phone / fax e-mail

Oral

X

Poster

Please mark your preference with X

Valderrama Valenzuela, Joaquin T. Department of signal theory, networking and communications. CITIC-UGR. University of Granada. Spain. C\ Periodista Daniel Saucedo Aranda s/n, 18004, Granada, Spain. Phone: +34 958 241 521, Fax: +34 958 240831. [email protected]

All messages will be sent ONLY to the corresponding author, please check very carefully your e-mail address Title of the abstract

A preliminary study of the short-term and long-term neural adaptation of the auditory brainstem response by the use of randomized stimulation

Authors

Valderrama J1, Alvarez I1, de la Torre A1, Segura JC1, Sainz M2,3, and Vargas J2

Affiliation of all Authors 1) Department of Signal Theory, Networking and Communications. CITIC(Institution/Hospital/Department, UGR. University of Granada. Spain. City, Country) 2) ENT Service. San Cecilio University Hospital. Granada. Spain. 3) Department of Surgery and Its Specialties. University of Granada. Spain. Before preparing your Abstract please read carefully the instructions given on the website

Submission deadline: February 16, 2012 Text only – no figures, no formulas. Maximum 300 words. TEXT: Brainstem auditory evoked response (BAER) signals represent the electrical activity of the auditory brainstem associated with a stimulus. The study of BAER at high stimulation rates is of great interest in the field of audiology since it presents several advantages: the reduction of the recording time, an earlier diagnosis of certain neural diseases, and the study of adaptation, which consists on a variation of the auditory response during a constant stimulus condition. This preliminary study is based on a novel stimulation technique that allows the recording of BAER at high rates of stimulation. This methodology consists on the average of auditory responses evoked by stimuli whose period varies randomly. Compared to other analogous techniques, this stimulation technique is the only methodology that allows the categorization of auditory responses according to the interval of the preceding stimulus. This premise has been used to design an experiment to check whether adaptation is a short-term or a long-term process. Only 6 normal hearing adults participated in this study. The results of this test suggest that though both factors are involved in the hearing process, subjects can be classified according to their tendency towards a short-term or a long-term adaptation process. Understand the biological mechanisms or the possible hearing diseases that influence such dispersion may have important repercussions in the field of audiology. Although a study with more subjects would be necessary to reach more solid conclusions, these preliminary results open up a new research line that may lead to a better understanding of the adaptation phenomenon.



A PRELIMINARY STUDY OF THE SHORT-TERM AND LONG-TERM NEURAL ADAPTATION OF THE AUDITORY BRAINSTEM RESPONSE BY THE USE OF RANDOMIZED STIMULATION Joaquín T. Valderrama, Isaac Álvarez, Ángel de la Torre, José Carlos Segura, Manuel Sáinz , José Luis Vargas

Department of Signal Theory, Networking and Communications. CITIC-UGR. University of Granada (Spain). ENT Service. San Cecilio University Hospital. Granada (Spain). Department of Surgery and its Specialties. University of Granada (Spain).

1 - INTRODUCTION Brainstem auditory evoked response (BAER) signals represent the electrical activity of the auditory brainstem associated with a stimulus. The study of BAER at high stimulation rates is of great interest in the field of audiology since it presents several advantages: the reduction of the recording time, an earlier diagnosis of certain neural diseases, and the study of adaptation, which consists on a variation of the auditory response during a constant stimulus condition. This preliminary study is based in BAER recorded at fast stimulation rates using the Randomized Stimulation technique to check wether adaptation is a short-term or a long-term process.

2 - RANDOMIZED STIMULATION & SPLIT The Randomized Stimulation technique consist of the average of auditory responses whose corresponding inter-stimulus interval (ISI) vary randomly between two values according to a predefined probability distribution (Figure 1.B, 1.C). This methodology can be used to obtain auditory evoked potentials at very fast stimulation rates (Figure 1.D). This stimulation technique allows the categorization of auditory responses according to the ISI of the preceding stimulus (Figure 2).

3 - EXPERIMENT BAER from six normally hearing subjects were recorded using the Randomized Stimulation technique at the stimulation rates ISI21-24 (rec), ISI2-5/21-24 (rec), and ISI2-5 (rec) (Figure 3). Auditory responses corresponding to ISI2-5 and ISI21-24 were retrieved from the ISI2-5/21-24 stimulation sequence, obtaining the ISI2-5 (split) and ISI21-24 (split) signals (Figure 3). Two scenarios were considered: (a) ISI21-24 (rec) and ISI2-5 (rec) signals are similar to ISI21-24 (split) and ISI2-5 (split) signals respectively; and (b) ISI21-24 (split) and ISI2-5 (split) are similar signals. The similarity of signals is analyzed in terms of amplitudes and latencies of the waves. On one hand, scenario (a) would suggest that adaptation is a short-term process since the morphology of the response is strongly influenced by the previous ISI. On the other hand, scenario (b) would indicate that adaptation is a long-term process in which the morphology of BAER depends on the stimulation rate of several previous stimuli. The results of this experiment show that amplitudes of the waves in ISI2-5 (split) signals are considerably smaller than those in ISI21-24 (split) signals in the six subjects. In contrast, the differences of Latency V between ISI2-5 (split) and ISI21-24 (split) signals vary among subjects: subjects 5 and 6 show a high difference, subjects 2 and 4 show a very small difference, and subjects 1 and 3 are in between. This findings suggest (1) that the mechanisms of adaptation that influence amplitudes and latencies in BAER are different; (2) that short-term and long-term adaptation mechanisms are involved in the hearing process; and (3) that subjects present a high dispersion according to their tendency towards a short-term or a long-term adaptation process. Understand the biological mechanisms or the possible hearing diseases that influence such dispersion may have important repercussions in the field of audiology. Although a study with more subjects would be necessary to reach more solid conclusions, these preliminary results let open a new research line that may lead to a better understanding of the adaptation phenomenon.

Figure 1. A) Stimulation sequence in the conventional technique. B) ISI3-8 Randomized Stimulation sequence. C) Histogram for an ISI3-8 Stimulation sequence. D) Examples of BAER recorded using the Randomized Stimulation technique at different stimulation rates.

Figure 3. BAER recorded from six subjects at the stimulation rates ISI21-24 (rec), ISI2-5/21-24 (rec), and ISI2-5 (rec); and BAER retrieved from the ISI2-5/21-24 sequence: ISI21-24 (split) and ISI2-5 (split). The number of sweeps recorded to obtain each signal is: ISI21-24 (rec) → 3000 sweeps; ISI2-5/21-24 (rec) → 10.000 sweeps; ISI21-24 (split) and ISI2-5 (split) → 5.000 sweeps; and ISI2-5 (rec) → 20.000 sweeps.

4 - CONCLUSIONS

Figure 2. Split & Average process explanation. A) ISI2-5/21-24 stimulation sequence. ISI of stimuli vary uniformly random between 2 - 5 and 21 - 24 ms. Auditory responses whose previous ISI is between 2 - 5 and 21 - 24 ms are highlighted in blue and red respectively. B) Histogram for a ISI2-5/21-24 stimulation sequence. C-D) Average of auditory responses whose previous ISI is between 21 - 24 ms (C) and between 2 - 5 (D).

E-mail us to [email protected]

✓ The proposed methodology can be used to explore the adaptation process. ✓ The mechanisms of adaptation that influence amplitudes and latencies in BAER are different. ✓ Both long-term and short-term adaptation mechanisms are involved in the hearing process. ✓ Subjects present a high dispersion in their tendency towards to a short-term or a long-term adaptation process. June 7-9, AHS Congress 2012, Cernobbio (Lake Como), Italy

2013 IEEE Point-of-Care Healthcare Technologies (PHT) Bangalore, India, 16 - 18 January, 2013

A portable, modular, and low cost auditory brainstem response recording system including an algorithm for automatic identification of responses suitable for hearing screening Joaquin T. Valderrama, Isaac Alvarez, Angel de la Torre, Jose C. Segura, Manuel Sainz, and Jose L. Vargas

Abstract— The recording of auditory brainstem response (ABR) signals is a common measure applied to assess hearing impairments. However, most of the available commercial devices able to record ABR signals can be unaffordable for many medical centers because of their cost and low flexibility. This paper describes a system that allows the recording of high quality ABR. Its low cost, easy handling, high performance, and portability make its use appropriate in low budget institutions. Furthermore, the flexibility and open nature of this system allow its use as a research tool. The ABR recording system includes a new algorithm for automatic evaluation of the quality of responses and the estimation of the latencies and amplitudes of the waves, the fitted parametric peaks (FPP). The performance of this technique is contrasted with a well-established method for quality evaluation based on the correlation coefficient. The encouraging results of this test suggest that the fitted parametric peaks could be used as a method for automatic ABR quality assessment and identification of the peaks.

and most of them give no access to raw recording data [4]. In contrast, the system described on this paper gives users full control of the parameters. Users are able to specify the intensity of stimulation, decide whether use the conventional stimulation technique or any other more advanced, select the number of biological responses to consider in the averaging process, set the stimulation frequency, define the analog to digital sample frequency, program the stimulation signal, change its polarity, set the filter settings, or implement advanced artifact rejection techniques. In addition, users have total access to raw recording data, which means that advanced digital processing can be implemented off-line. Furthermore, the low cost nature of this system and its high performance allow a reliable use in many low-budget institutions.

I. INTRODUCTION Auditory brainstem response (ABR) signals represent the electrical activity of the auditory nerve associated to a stimulus. This biological response is characterized by waves (peaks) that occur whithin the first 10 ms post-stimulus interval [1]. Peaks are labelled by roman numerals as proposed Jewett in [2]. ABR signals are widely used in hospitals and clinics around the world as a hearing screening method to detect hearing threshold and hearing impairments [1], [3]. Moreover, the study of auditory evoked potentials is of great interest in audiology since it allows the analysis of the mechanisms involved in the process of hearing [1]. This paper describes a high performance, portable, modular, and low cost auditory brainstem response recording system. There already exist several commercial devices able to record ABR; nevertheless most of the current clinical systems only allow users to select a few parameter settings, are expensive, require connection to the electrical network, This research is granted by the project “Design, implementation and evaluation of an advanced system for recording Auditory Brainstem Response (ABR) based in encoded signalling” (TEC2009-14245), R&D National Plan (2008-2011), Ministry of Economy and Competitivity (Government of Spain) and “European Regional Development fund Programme” (20072013); and by the grant “Universitary Professor Training Program” (FPU, AP2009-3150), Ministry of Education, Culture, and Sports (Government of Spain). J. T. Valderrama, I. Alvarez, A. de la Torre, and J. C. Segura are with the Department of Signal Theory, Telematics and Communications. CITICUGR. University of Granada. Spain. (e-mail: [email protected]). M. Sainz and J. L. Vargas are with San Cecilio University Hospital. ENT Service. Granada. Spain M. Sainz is with the Department of Surgery and its Specialties. University of Granada. Spain.

978-1-4673-2767-1/13/$31.00 ©2013 IEEE

The flexibility provided by the ABR recording system described on this paper allows its use as a research tool. The ABR recording system incorporates a new software module that provides an automatic evaluation of the quality of ABR signals based on the use of templates. This methodology is called fitted parametric peaks (FPP). The automatic analysis of the quality of ABR recordings can be useful to take the decision of automatically stop averaging, avoiding the recording of unnecessary responses when there already exists an ABR signal of enough quality and therefore, reducing recording time [5]. It also allows an automated identification of the parameters of the peaks, i.e., amplitudes and latencies [6], which can be used to provide an automated interpretation of the auditory brainstem response [7]. Besides, automated algorithms remove the need for subjective interpretations of ABR, reducing human errors, and ensuring consistency among patients, test conditions, and screening personnel [8]. These advantages led the Joint Committee on Infant Hearing Year 2000 Position Statement to declare: “screening technologies that incorporate automated response detection are preferred over those that require operator interpretation and decision making” [9]. The fitted parametric peaks methodology is validated in this study contrasting its performance with a well-established method for quality assessment of ABR recordings based on the correlation coefficient [10]. The main advantages of the proposed methodology for automated ABR quality assessment are discussed on this paper.

180

Fig. 1.

General scheme of the system. Fig. 2. Parameters involved in the SNR estimation based on fitted parametric peaks. The parametric peak fitted to the wave V of a test signal is shown in bold.

II. METHODS A. System description ABR signals are recorded conventionally through the presentation of a stimulus that excites the auditory system of a subject and the recording of its associated auditory electrical response. This biological response is collected by surface electrodes placed on the skin at different positions on the head. The low amplitude of this kind of signal (usually less than 1µV ) forces to make a high amplification. Evoked responses are contaminated by several sources of artifacts such us neuro-muscular activity of the subject, noise associated with the amplifier and electromagnetic & radiofrequency interferences. The methodology used to reduce the effects of these artifacts is the average of a large number of biological responses in order to improve the signal to noise ratio [11]. This system is battery powered in order to minimize the artifact produced by the electric network. Signal processing has been developed with MATLAB (The Mathworks, Inc.). The process of ABR recording is sketched on figure 1. A specific signal composed of a sequence of a 0.1 ms duration clicks is generated by the laptop for both stimulation and synchronization purposes. The open nature of this system allows the use of any kind of stimulation sequence: conventional; MLS [12], [13]; CLAD [14], [15]; QSD [16]; or randomized stimulation [17]. These techniques allow the recording of ABR at different stimulation rates, which can be useful in some applications such as reducing the recording time [18], or detecting certain pathologies at an early stage [19]. The intensity of the stimulation can also be controlled by setting the amplitude of the clicks. The stimulation and synchronization signal is sent synchronously through the left and right outputs of an external Analog-Digital/DigitalAnalog (AD/DA) sound card. The right output of the AD/DA sound card is connected to its left input, so the recording of the synchronization signal allows the system to determine the exact moment in which stimuli are produced. The left output of the sound card connects a pair of headphones, through which the stimulation signal excites the auditory system of a subject. The auditory electrical response associated to each stimulus is recorded by three Ag/AgCl surface disc electrodes. The electroencephalogram (EEG) captured by the electrodes is preamplified by a factor G1 = 25, band-pass filtered (150-3000 Hz), and amplified by a factor G2 = 130. Therefore, the gain of the amplifier for the band-pass

frequencies is set at about Gamp = 3250 (70 dB). The auditory response after filtering and amplification, and the synchronization signal are recorded synchronously by the right and left inputs of the external AD/DA sound card. Both signals are sampled at a frequency of 25 kHz and stored using 16 bits of quantization. Finally, the auditory evoked response is obtained through the average of the auditory responses. The rough cost of this ABR recording system (laptop not included) is lower than $500. A full description of the ABR recording system can be browsed in [20]. B. Quality assessment of ABR responses The described ABR recording system incorporates a new approach for automatic assessment of quality of ABR recordings. We have called this quality evaluation technique fitted parametric peaks (FPP). This new procedure is based on the use of templates that fit the ABR response, and are used as reference to evaluate the quality of the ABR signal. The use of templates for this purpose is a well-established method [6]. The fitted parametric peaks methodology uses as template the following function: (t−L)2 (t − L)2 x(t, A, L, W ) = A · 1 − · e− 2·W 2 2 W

(1)

This parametric function corresponds to the second derivative of a Gaussian function of amplitude A, mean L, and standard deviation W . In this parametric peak, L represents the latency, W is the semi-width, and A is the amplitude of the peak (figure 2). Given an auditory brainstem response signal, the parameters A, L, and W can be iteratively estimated with a criterion of minimum mean square error. In order to evaluate the quality of an ABR recording, the reference parameters (Aref , Lref , and Wref ) can be obtained for waves III and V according to previous literature. The reference parameters considered in this study are the followings: Aref = 0.26µV for wave III and Aref = 0.28µV for wave V; Lref = 3.75ms for wave III and Lref = 5.80ms for wave V, and Wref = 0.4ms for both waves [1]. The parameters (Atest , Ltest , and Wtest ) for waves III and V are then estimated in the test signal in intervals around the reference parameters: Ltest ∈ Lref ± 0.5 ms, Wtest ∈ [0.4 · Wref , 2 · Wref ], Atest ∈ [0.2 · Aref , 4 · Aref ]. The SNR can be

181

Fig. 3. A) Examples of ABR signals from 4 subjects: ISI=22ms, 70 dBnHL, 1000 auditory responses. The quality evaluation of these recordings by both methodologies is presented. B) Evaluation of the quality of ABR recordings obtained at different number of averaged responses according to the fitted parametric peaks (FPP) and to the correlation coefficient (r) methodologies. C) Relationship of the quality of ABR signals assessed by the fitted parametric peaks and by the correlation coefficient techniques. The averaged experimental data are adjusted to a first order model. The dashed line represents the standard deviation of the experimental data.

estimated independently for each wave as the ratio between the pewer of the parametric peak and the pewer of the noise (estimated as the difference between the ABR response and the parametric peak). The SNR for each peak is estimated in an interval [L-4W L+4W ]. This way, the evaluation of an ABR recording does not degrade because of small fluctuations in the amplitude, latency or width with respect to the reference signal. This criterion for the evaluation of the response approaches the subjective evaluation provided by an experienced audiologist, essentially based on the grade of identification of the most important waves. The performance of the proposed fitted parametric peaks (FPP) methodology is compared to a standard ABR quality assessment based on the correlation coefficient (r). This parameter points out the grade of similarity between two ABR responses. A high possitive correlation coefficient would indicate a high quality ABR if both signals are recorded in similar conditions [21]. Compared to other automatic quality evaluation techniques, the correlation coefficient remains as the most consistent technique to score the quality of ABR recordings [10]. The performance of the fitted parametric peaks methodology is contrasted in this paper with the quality assessment technique based on the correlation coefficient. 20 000 auditory responses were recorded from 4 normally hearing adults at an intensity of 70 dBnHL, and at the stimulation rate 45,45 Hz − Interstimulus interval (ISI) = 22 ms. All available auditory responses were grouped in blocks of a specific number of responses. The number of responses of each block varied from 500 to 9500. For instance, there were 40 blocks of 500 auditory responses, 20 blocks of 1000 responses, and 2 blocks of 9500 responses in each subject. The quality of the ABR signal obtained in each block was evaluated using the fitted parametric peaks methodology.

Then, all evaluations corresponding to blocks of the same number of responses were averaged to obtain a final quality evaluation based in FPP. The quality evaluation based on the correlation coefficient was performed by calculating the correlation coefficient between all possible combinations of ABR corresponding to blocks of the same number of responses and same subject, taking two at a time; and finally, averaging all these evaluations. The quality evaluation technique based on fitted parametric peaks was validated by contrasting its performance with the methodology based on the correlation coefficient. The functional dependence between both methodologies was analysed making groups of 0.2 dB from quality evaluations provided by the FPP, and calculating the mean and standard deviation of the corresponding quality evaluations based on the correlation coefficient. III. RESULTS Figure 3.A shews two examples per subject of ABR signals obtained using 1000 auditory responses in the averaging process. The most important waves can be identified in all subjects, in exception of subject four, that wave I cannot be clearly idenfied. Waves I, III, and V are labelled in the figure. This figure also remarks differences in the morphology among subjects. All subjects present similar amplitudes and latencies, which are consistent with previous literature [6], [7], [1]. The evaluation of the quality of these recordings by the fitted parametric peaks (FPP) and by the correlation coefficient (r) methodologies is presented in the figure. Figure 3.B presents the evaluation of the quality of ABR recordings obtained at an increasing number of averaged responses by both the FPP and by the correlation coefficient methodologies. This figure shews that the quality of the responses increases with the number of averaged responses

182

in both techniques. The ABR recording system mentioned in this article allows the recording of high quality ABR using a fair number of auditory responses in the averaging process (r ≈ 0.9 at 1000 averaged responses). The evaluation of quality provided by the FPP methodology is compared to the evaluation given by the correlation coefficient technique (figure 3.C). Despite both methodologies assess quality by different means, this graphic shows that both techniques present in average a similar tendency. This tendency is characterized by a linear regression analysis of the experimental data (r = 0.88). Consequently, both techniques could be used to provide automatic evaluation of quality of ABR recordings. Moreover, a subjective evaluation given by an experienced audiologist suggest that the quality evaluation method FPP is more consistent because the correlation coefficient could provide an inaccurate high evaluation when comparing two ABR recordings contaminated, i.e., by a strong stimulus artifact. IV. DISCUSSION AND CONCLUSION This paper describes an ABR recording system. The high performance and low-price of this system could make its use appropriate in low-budget institutions and medical centers from developing countries. The high portability provided by this battery powered system could also spread a hearing screening protocol to rural and other difficult access areas by driving the device to such areas, instead of restraining its use in medical centers. In contrast to many commercial devices, the proposed ABR recording system could also be used as a research platform due to its open nature and flexibility. The described ABR recording system includes a software module that provides an automatic evaluation of the quality for an ABR recording. This technique is called fitted parametric peaks (FPP). The performance of this technique was contrasted with the well-established quality evaluation method based on the correlation coefficient. In comparison with this technique, the FPP methodology: (a) requires only one ABR recording to provide a quality estimation, which can be useful in applications where clinical test time is a critical parameter, such as in ongoing quality assessment applications; (b) is more consistent in certain situations such as evaluating the quality in recordings contaminated by a large stimulus artifact, in which an evaluation of the quality based on the correlation coefficient would result into an inaccurate high quality evaluation; and (c) provides an automatic identification of the amplitude and latency of the peaks. The preliminary results presented on this paper suggest (a) that the ABR recording system presented in this paper can be used to obtain reliable and high-quality ABR recordings, and (b) that the proposed FPP can be considered a valid procedure to provide an automatic assessment of the quality of ABR recordings and identification of the peaks. ACKNOWLEDGMENT Authors acknowledge the collaboration of the subjects that have participated in this study.

R EFERENCES [1] J. W. Hall, “New handbook of Auditory Evoked Responses,” Ed. Pearson: Allyn and Bacon (Boston), 2007, pp. 750. [2] D. L. Jewett and J. S. Williston, “Auditory evoked far fields averaged from the scalp of humans,” Brain, vol. 94, no. 4, pp. 681-696, 1971. [3] A. Erenberg, J. Lemons, C. Sia, D. Tunkel, P. Ziring, M. Adams, J. Hostrum, M. McPherson, N. Paneth, and B. Strickland, “Newborn and infant hearing loss: Detection and intervention,” Pediatrics, vol. 103, no. 2, pp. 527-530, 1999. [4] A. Bahmer, O. Peter and U. Baumann, “Recording of electrically evoked auditory brainstem responses (E-ABR) with an integrated stimulus generator in Matlab,” Journal of Neuroscience Methods, vol. 173, no. 2, pp. 306-314, 2008. [5] O. Ozdamar and R. E. Delgado, “Measurement of signal and noise characteristics in ongoing auditory brainstem response averaging,” Annals of Biomedical Engineering, vol. 24, no. 6, pp. 702-715, 1996. [6] C. Elberling, “Auditory Electrophysiology: The use of templates and cross correlation functions in the analysis of brain stem potentials,” Scandinavian Audiology, vol. 8, no. 3, pp. 187-190, 1979. [7] R. E. Delgado and O. Ozdamar, “Automated auditory brainstem response interpretation,” IEEE Engineering in Medicine and Biology, vol. 13, no. 2, pp. 227-237, 1994. [8] G. G. Gentiletti-Faenze, O. Yanez-Suarez, and J. M. Cornejo-Cruz, “Evaluation of automatic identification algorithms for auditory brainstem response used in universal hearing loss screening,” Proceedings of the 25th Annual International Conference of the IEEE EMBS Proceedings 3, pp. 2857-2860, Cancun, Mexico, September 17-21, 2003. [9] Joint Committee on Infant Hearing Joint Year 2000 Position Statement, “Principles and Guidelines for Early Hearing Detection and Intervention Programs,” 2000. [10] S. A. Arnold, “Objective versus visual detection of the auditory brain stem response,” Ear and Hearing, vol. 6, no. 3, pp. 144-150, 1985. [11] P. K. H. Wong and R. G. Bickford, “Brain stem auditory evoked potentials: the use of noise estimate,” Electroencephalography and Clinical Neurophysiology, vol. 50, no. 1-2, pp. 25-34, 1980. [12] U. Eysholdt and C. Schreiner, “Maximum length sequences - A fast method for measuring brainstem auditory evoked responses,” 3rd Annual Conference of the IEEE EMBS, Houston, TX, USA, pp. 306309, 1981. [13] U. Eysholdt and C. Schreiner, “Maximum length sequences - A fast method for measuring brain-stem-evoked responses,” Audiology, vol. 21, no. 3, pp. 242-250, 1982. [14] O. Ozdamar, R. E. Delgado, E. Yavuz, K. Thombre, and N. Acikgoz, “Deconvolution of auditory evoked potentials obtained at high stimulus rates,” Proc. 1st Int. IEEE EMBS Conf. Neural Engineering, Capri, Italy, pp. 285-288, 2003. [15] R. E. Delgado and O. Ozdamar, “Deconvolution of evoked reponses obtained at high stimulus rates,” Journal of the Acoustical Society of America, vol. 115, no. 3, pp. 1242-1251, 2004. [16] D. L. Jewett, G. Caplovitz, B. Baird, M. Trumpis, M. P. Olson, and L. J. Larson-Prior, “The use of QSD (q-sequence deconvolution) to recover superposed, transient evoked-responses,” Clinical Neurophysiology, vol. 115, no. 12, 2754-2775, 2004. [17] I. Alvarez, J. T. Valderrama, A. DeLaTorre, J. C. Segura, M. Sainz, and J. L. Vargas, “Reducción del tiempo de exploración de potenciales evocados auditivos del tronco cerebral mediante estimulación aleatorizada” - [“Reduction of recording time of auditory brainstem response by the use of randomized stimulation”], in XXV Simposium URSI, 2010. [18] S. Leung, A. Slaven, A. R. D. Thornton, and G. J. Brickley, “The use of high stimulus rate auditory brainstem responses in the estimation of hearing threshold,” Hearing Research, vol. 123, no. 1-2, pp. 201-205, 1998. [19] Z. D. Jiang, D. M. Brosi, X. M. Shao, and A. R. Wilkinson, “Maximum length sequence brainstem auditory evoked responses in term neonates who have perinatal hypoxia-ischemia,” Pediatric Research, vol. 48, no. 5, pp. 639-645, 2000. [20] J. T. Valderrama, I. Alvarez, A. DeLaTorre, J. C. Segura, M. Sainz, J. L. Vargas, “Educational approach of a BAER recording system based on experiential learning,” Technics Technologies Education Management, vol. 6, no. 4, pp. 876-889, 2011. [21] S. M. Mason, A. P. Su, and R. A. Hayes, “Simple online detector of auditory evoked cortical potentials,” Medical and Biological Engineering and Computing, vol. 15, no. 6, pp. 641-647, 1977.

183

04/08/2014

A portable, modular, and low cost auditory brainstem response recording system including an algorithm for automatic identification of responses suitable for hearing screening Joaquin T. Valderrama*, Isaac Alvarez, Angel de la Torre, Jose C. Segura, Manuel Sainz, and Jose L. Vargas

*E‐mail us to [email protected]

Bangalore (India), 16‐18 January 2013

IEEE EMBS Special Topic Conference on Point‐of‐Care Healthcare Technologies

Introduction. ABR signals • The human auditory system

• Clinical application Source: SPD Australia IEEE EMBS Special Topic Conference on Point‐of‐Care Healthcare Technologies 2 / 8

Bangalore (India), 16‐18 January 2013

1

04/08/2014

Introduction. ABR recording process • Recording problems • Amplitude

Development and implementation of an advanced

Development and implementation of an advanced

Suggest Documents

Development of an Implementation Intervention

Development and Implementation of an Extensible ...

The development and implementation of an ...

development, integration and implementation of an ...

Development and Implementation of an Impedance ...

development, implementation and testing of an ...

Development and Classroom Implementation of an Environmental

Implementation of an Advanced Nursing Directive for

Factors influencing the development and implementation of advanced ...

Design, implementation, and on-sky performance of an advanced ...

An Advanced Computer Algorithm for Implementation ...

Development of an implementation intention ... - Semantic Scholar

Development of an Advanced Planetary Space Suit

Development of an advanced web application for

Development of an advanced exposure assessment ...

Development of an Advanced Deformable ... - Semantic Scholar

Development of an Advanced Rotorcraft Preliminary ... - CiteSeerX

Development and implementation of an herbal and natural product ...

Design, Implementation and Testing of Advanced ...

Design, Implementation and Testing of Advanced

Design and Implementation of Advanced Algorithms

Design, Implementation and Testing of Advanced ...

An analysis of the development and implementation of a ...

Development and Implementation of C4NP