Wavelet transform feature extraction from human ... - Semantic Scholar

Digital Signal Processing 18 (2008) 861–874 www.elsevier.com/locate/dsp

Wavelet transform feature extraction from human PPG, ECG, and EEG signal responses to ELF PEMF exposures: A pilot study Dean Cvetkovic a , Elif Derya Übeyli b,∗ , Irena Cosic a a RMIT University, School of Electrical and Computer Engineering, GPO Box 2476V, Melbourne, Victoria 3001, Australia b TOBB Economics and Technology University, Faculty of Engineering, Department of Electrical and Electronics Engineering, 06530 Sö˘gütözü,

Ankara, Turkey Available online 6 June 2007

Abstract This paper presents the experimental pilot study to investigate the effects of pulsed electromagnetic field (PEMF) at extremely low frequency (ELF) in response to photoplethysmographic (PPG), electrocardiographic (ECG), electroencephalographic (EEG) activity. The assessment of wavelet transform (WT) as a feature extraction method was used in representing the electrophysiological signals. Considering that classification is often more accurate when the pattern is simplified through representation by important features, the feature extraction and selection play an important role in classifying systems such as neural networks. The PPG, ECG, EEG signals were decomposed into time-frequency representations using discrete wavelet transform (DWT) and the statistical features were calculated to depict their distribution. Our pilot study investigation for any possible electrophysiological activity alterations due to ELF PEMF exposure, was evaluated by the efficiency of DWT as a feature extraction method in representing the signals. As a result, this feature extraction has been justified as a feasible method. © 2007 Elsevier Inc. All rights reserved. Keywords: Feature extraction; PPG; ECG; EEG; Discrete wavelet transform; PEMF; ELF; Bioeffects

1. Introduction 1.1. Feature extraction A feature is a distinctive or characteristic measurement, transform, structural component extracted from a segment of a pattern. Features are used to represent patterns with the goal of minimizing the loss of important information. The feature vector, which is composed of the set of all features used to describe a pattern, reduces the dimensional space needed to represent that pattern. This, in effect, means that the set of all features that could be used to describe a given pattern (a large and in theory an infinite number as very small changes in some parameters are allowed to separate different features) is limited to those actually represented in the feature vector. One purpose of the dimension reduction is satifying engineering constraints in software and hardware complexity, the cost of data processing, and * Corresponding author. Fax: +90 312 2924091.

E-mail address: [email protected] (E.D. Übeyli). 1051-2004/$ – see front matter © 2007 Elsevier Inc. All rights reserved. doi:10.1016/j.dsp.2007.05.009

862

D. Cvetkovic et al. / Digital Signal Processing 18 (2008) 861–874

the (why else?) desirability of compressing pattern information. In addition, the classification often becomes more accurate when the pattern is simplified by including important features or properties only [1–3]. A feature extraction is the determination of a feature or a feature vector from a pattern vector. In order to make pattern processing problems solvable one needs to convert patterns into features, which become condensed representations of patterns, ideally containing only salient information. Feature extraction methods could be based on either calculating statistical characteristics or producing syntactic descriptions. The feature selection process usually is designed to provide a means for choosing the features which are best for classification optimized against on various criteria. The feature selection process performed on a set of predetermined features. Features are selected based on either (1) best representation of a given class of signals, or (2) best distinction between classes. Therefore, feature selection plays an important role in classifying systems such as neural networks. For the purpose of classification problems, the classifying system has usually been implemented with rules using ifthen clauses, which state the conditions of certain attributes and resulting rules. However, it has proven to be a difficult and time consuming method. From the viewpoint of managing large quantities of data, it would still be most useful if irrelevant or redundant attributes could be segregated from relevant and important ones, although the exact governing rules may not be known. In this case, the process of extracting useful information from a large data set can be greatly facilitated [1–3]. Various methodologies of automated diagnosis have been adopted, however the entire process can generally be subdivided into a number of disjoint processing modules: preprocessing, feature extraction/selection, and classification. The signal acquisition, artefact removing, averaging, thresholding, signal enhancement, and edge detection are the main operations in the course of preprocessing. The accuracy of signal/image acquisition is of great importance since it contributes significantly to the overall classification result. The markers are subsequently processed by the feature extraction module. The module of feature selection is an optional stage, whereby the feature vector is reduced in size including only, from the classification viewpoint, what may be considered as the most relevant features required for discrimination. The classification module is the final stage in automated diagnosis. It examines the input feature vector and based on its algorithmic nature, produces a suggestive hypothesis [3–5]. In the feature extraction stage, numerous different methods can be used so that several diverse features can be extracted from the same raw data. The wavelet transform (WT) provides very general techniques which can be applied to many tasks in signal processing. Wavelets are ideally suited for the analysis of sudden short-duration signal changes. One very important application is the ability to compute and manipulate data in compressed parameters which are often called features [6]. Thus, the time-varying biomedical signal, consisting of many data points, can be compressed into a few parameters by the usage of the WT. These parameters characterize the behavior of the time-varying biomedical signal. This feature of using a smaller number of parameters to represent the time-varying biomedical signal is particularly important for recognition and diagnostic purposes [7–13]. The objective of the present study in the field of automated detection of changes in time-varying biomedical signals is to extract the representative features of the signals under study in order to obtain the accurate classification models. 1.2. ELF PEMF bioeffects In the past few decades, the responses of human and animal electrophysiological signal activity to nonionising extremely low frequency (ELF) pulsed electromagnetic field (PEMF) radiation have been studied [14–25]. Since that time, various studies have reported that humans and animals are particularly sensitive to ELF or ELF modulated sensory stimulation. The ELF refers to the range of electromagnetic field frequencies below 300 Hz. Several studies have examined the effects of sinusoidal ELF magnetic fields upon the human EEG activity in the past [15]. Cvetkovic et al. single-blind counter-balanced pilot study investigated whether the human EEG activity could be altered when stimulated by localised ELF magnetic field at the top-central human head region [16]. This pilot study (8 subjects) revealed a marginal significant decrease in exposure compared to control, found in the Alpha1 EEG band (7.5–9.5 Hz) at the vertex head position, where magnetic field stimulation was applied at the Alpha1 frequency of 174 µT/8.33 Hz. It has been assumed that the effect in Alpha1 EEG findings is possibly related to ‘synchronisation,’ ‘induced rhythmic,’ and ‘synchrony spread’ theories of neuron firing rate after ELF magnetic field. Bell et al. study on 20 subjects, exposed to 60 Hz and 25–50 µT, found an increase in the spectral power (all EEG bands) at the central and the parietal regions [17]. Similar study by Bell et al. reported a decrease in the Alpha (8–13 Hz) EEG activity at the occipital region after 2 s magnetic field exposure of 10 Hz, 100 µT and an increase


863

in the Alpha 10 Hz EEG activity at the central region after 10 min magnetic field exposure of 10 Hz/40 µT and 1.5 Hz/20 µT [18]. Marino et al. study found increases in the spectral power mainly at higher than 10 Hz EEG frequencies at the central, parietal and occipital regions (10 and 1.5 Hz, 80 µT) after 2 s magnetic field exposures [19]. Hausser et al. found significant increases in the Theta (3.5–7.5 Hz) and the Beta (12.5–25 Hz) EEG bands at the occipital region after 20 min magnetic field exposure of 3 Hz and 100 µT [20]. A study by Lyskov et al. found a decrease in the Delta (1–3 Hz) and the Theta (4–7 Hz) spectral power at the frontal, central and parietal regions and an increase in the Alpha (7–13 Hz) power at the occipital and the Beta (14–25 Hz) in the frontal regions after a 60 min intermittent (1 s on and off) magnetic field exposure of 45 Hz and 1.26 mT [21]. Ghione et al. study on 40 human subjects indicated that the 90 min exposure of ELF magnetic fields (50 Hz, 80 µT) could influence a significant increase in the Alpha (8–13 Hz) EEG activity [22]. Sestre et al. study on 50 human subjects examined the effects of controlled changes in the Earth’s (DC) magnetic field on EEG. They found a significant increase in the Delta and Theta EEG bands observed in the controlled and artificial North Pole conditions (45–57.4 µT) which were generated by tri-axil Helmholtz configuration [23]. Tabor et al. study on 15 subjects revealed that the changes of time-domain heart rate variability (HRV) parameters could be associated with the influence of 50 Hz magnetic field (20–30 µT) [24]. Tabor’s time-domain HRV parameters included linear and nonlinear analysis. Baldi et al. study on the influence of ELF PEMF exposure on the HRV using linear analysis, revealed a LF/HF variation [25]. In comparison to previous studies, the pilot study described in this paper consists of measuring the EEG, ECG, and PPG signal responses to ELF PEMF exposures over a 5 day period. However, our paper’s main contribution to bioelectromagnetism and digital signal processing body of knowledge is the employment of an unique signal processing method using wavelet feature extraction of EEG, ECG and PPG signals in order to examine the effects of ELF PEMF upon the human body. 2. Data description The pilot experiment consisted of one healthy subject (gender: male, age: 25 years), recruited to participate for 5 consecutive days (excluding weekends). The RMIT ethics committee approved the study and all subjects gave written informed consent prior to the experiment. The experiment was a double-blinding counter-balanced design. There were total of 25 recording sessions. At each sessions (day) the experimental protocol was designed to record the biosignals before (baseline) and after (post) EMF exposure during: control (no magnetic field) and exposure (magnetic field). The ELF PEMF stimulations were undertaken with Bioresonance BRS-500 system (Medec Limited, Australia). The BRS magnetic field exposure system consists of the programmed control unit, an applicator matters and pillow. The mattress (60 × 140 cm in size) was used as a magnetic field exposure system, consisted of 3 coil pairs, embedded within the mattress and spaced strategically in order to apply the exposure to the whole body area. The 10 min magnetic field exposure, in the direction perpendicular to the body and mattress was measured to be 8.3 µT (subject’s back of the neck), 2 µT (chest region), and 1 µT (below chest region). The signal generated from the control unit to create PEMF, consisted of four identical ‘saw-tooth’ waveforms, 0–100 Hz frequency ‘bundles’ and delays, each lasting 20 ms. All the biosignals were recorded for 60 s after the ELF PEMF exposure. In order to shield the subjects and electrodes from the 50 Hz electric fields in the laboratory, the recording took place in a dim room, inside a Faraday cage (1.95 × 1.83 × 2.67 m) constructed of mesh wire (2.5 × 2.5 cm) and steel frames. Inside the cage, subjects were laid in a comfortable semi-reclining Metron chair. The lead-I electrocardiographic (ECG) and photoplethysmographic (PPG) signals (sampled at 250 Hz) were recorded using Biopac Inc. data acquisition device which consisted of MP100A system with ECG100C and PPG100C amplifiers, respectively. The analog 50 Hz notch filter was activated to filter any artefacts in the recorded signals from the main sources. The signals were transmitted to a PC’s Acknowledge 3.7 software via USB cable. The recordings of the electroencephalographic (EEG) activity were conducted using Mindset EEG system together with standard Neuroscan 19-electrode EEG cap. The cap was placed on subject’s head according to 10/20 International System. The Referential Montage of 16 channels was used throughout this investigation. The left brain hemisphere electrodes: Fp1, F7, F3, T7, C3, P7, P3, and O1 were all referenced to M1 or A1 (left mastoid). While the right brain hemisphere electrodes: Fp2, F8, F4, T8, C4, P8, P4, and O2 were referenced to right mastoid (M2 or A2). The EEG signals were sampled at the rate of 256 Hz.

864


3. Wavelet transform The WT is designed to address the problem of nonstationary signals. It involves representing a time function in terms of simple, fixed building blocks, termed wavelets. These building blocks are actually a family of functions which are derived from a single generating function called the mother wavelet by translation and dilation operations. Dilation, also known as scaling, compresses or stretches the mother wavelet and translation shifts it along the time axis [3,6,8,10]. The WT can be categorized into continuous and discrete. Continuous wavelet transform (CWT) is defined by +∞ ∗ x(t)ψa,b (t) dt, CWT(a, b) =

(1)

−∞

where x(t) represents the analyzed signal, a and b represent the scaling factor (dilatation/compression coefficient) and translation along the time axis (shifting coefficient), respectively, and the superscript asterisk denotes the complex conjugation. ψa,b (·) is obtained by scaling the wavelet at time b and scale a: t −b 1 , (2) ψa,b (t) = √ ψ a |a| where ψ(t) represents the wavelet [6,8]. Continuous, in the context of the WT, implies that the scaling and translation parameters a and b change continuously. However, calculating wavelet coefficients for every possible scale can represent a considerable effort and result in a vast amount of data. Therefore discrete wavelet transform (DWT) is often used. The WT can be thought of as an extension of the classic Fourier transform, except that, instead of working on a single scale (time or frequency), it works on a multi-scale basis. This multi-scale feature of the WT allows the decomposition of a signal into a number of scales, each scale representing a particular coarseness of the signal under study. The procedure of multiresolution decomposition of a signal x[n] is schematically shown in Fig. 1. Each stage of this scheme consists of two digital filters and two downsamplers by 2. The first filter, g[·] is the discrete mother wavelet, high-pass in nature, and the second, h[·] is its mirror version, low-pass in nature. The downsampled outputs of first high-pass and low-pass filters provide the detail, D1 and the approximation, A1 , respectively. The first approximation, A1 is further decomposed and this process is continued as shown in Fig. 1. All wavelet transforms can be specified in terms of a low-pass filter h, which satisfies the standard quadrature mirror filter condition: H (z)H (z−1 ) + H (−z)H (−z−1 ) = 1,

(3)

where H (z) denotes the z-transform of the filter h. Its complementary high-pass filter can be defined as G(z) = zH (−z−1 ).

Fig. 1. Subband decomposition of discrete wavelet transform implementation; g[n] is the high-pass filter, h[n] is the low-pass filter.

(4)


865

A sequence of filters with increasing length (indexed by i) can be obtained: i

Hi+1 (z) = H (z2 )Hi (z),

i

Gi+1 (z) = G(z2 )Hi (z),

i = 0, . . . , I − 1

(5)

with the initial condition H0 (z) = 1. It is expressed as a two-scale relation in time domain hi+1 (k) = [h]↑2i ∗ hi (k),

gi+1 (k) = [g]↑2i ∗ hi (k),

(6)

where the subscript [·]↑m indicates the up-sampling by a factor of m and k is the equally sampled discrete time. The normalized wavelet and scale basis functions ϕi,l (k), ψi,l (k) can be defined as ϕi,l (k) = 2i/2 hi (k − 2i l),

ψi,l (k) = 2i/2 gi (k − 2i l),

(7)

where the factor 2i/2 is an inner product normalization, i and l are the scale parameter and the translation parameter, respectively. The DWT decomposition can be described as a(i) (l) = x(k) ∗ ϕi,l (k),

d(i) (l) = x(k) ∗ ψi,l (k),

(8)

where a(i) (l) and di (l) are the approximation coefficients and the detail coefficients at resolution i, respectively [6,8]. The concept of being able to decompose a signal totally and then perfectly reconstruct the signal again is practical, but it is not particularly useful by itself. In order to make use of this tool it is necessary to manipulate the wavelet coefficients to identify characteristics of the signal that were not apparent from the original time domain signal. 4. Results 4.1. Feature extraction using discrete wavelet transform The spectral analysis of the PPG, ECG, EEG signals was performed using the DWT as described in Section 3. The selection of appropriate wavelet and the number of decomposition levels is very important in analysis of signals using the WT. The number of decomposition levels is chosen based on the dominant frequency components of the signal. The levels are chosen such that those parts of the signal that correlate well with the frequencies required for classification of the signal are retained in the wavelet coefficients. In the present study, the number of decomposition levels was chosen to be 4. Thus, the PPG, ECG, EEG signals were decomposed into the details D1 –D4 and one final approximation, A4 . Usually, tests are performed with different types of wavelets and the one which gives maximum efficiency is selected for the particular application. The smoothing feature of the Daubechies wavelet of order 2 (db2) made it more suitable to detect changes of the signals under study. Therefore, the wavelet coefficients were computed using the db2 in the present study. The frequency bands corresponding to different levels of decomposition for Daubechies wavelet of order 2 (db2) with a sampling frequency of 256 Hz are: D1 (37.5–75 Hz); D2 (18.75– 37.5 Hz); D3 (9.375–18.75 Hz); D4 (4.6875–9.375 Hz); and A4 (0–4.6875 Hz). The discrete wavelet coefficients were computed using the MATLAB software tool. The feature selection is an important component of designing the neural network based on pattern classification since even the best classifier will perform poorly if the features used as inputs are not selected well. The computed discrete wavelet coefficients provide a compact representation that shows the energy distribution of the signal in time and frequency. Therefore, the computed discrete wavelet coefficients of the PPG, ECG, EEG signals of each record were used as the feature vectors representing the signals. In order to reduce the dimensionality of the extracted feature vectors, statistics over the set of the wavelet coefficients was used. The following statistical features were used to represent the time-frequency distribution of the signals under study: 1. 2. 3. 4.

Maximum of the wavelet coefficients in each subband. Mean of the wavelet coefficients in each subband. Minimum of the wavelet coefficients in each subband. Standard deviation of the wavelet coefficients in each subband.

There were 45 recordings in total performed for EEG, ECG, and PPG. For the ECG, PPG, and EEG recordings, as described in Section 2, there were 16 channels (EEG) and 2 channels (ECG and PPG) recorded. Due to a large number of factors needed to be processed and analysed, such as wavelet coefficients of 5 subbands, 16 channels, 5 days and

866


(a) Fig. 2. Extracted (a) maximum, (b) mean, (c) minimum, and (d) standard deviation statistical features of exemplary PPG and ECG recordings. The statistical feature of the wavelet coefficients in each subband is representated. The figure’s labels are as follows: x-axis represent the days of recordings; y-axis—ratio (after/before); legend—control/exposure; top rows—ECG; bottom rows—PPG; and columns represent wavelet coefficients in each subband.

pre-exposure(before)/post-exposure(control/exposure) conditions, it was necessary to employ a data reduction. This data reduction included the averaging of 16 EEG channels into 4 main brain regions (i.e., frontal, central, pariatel, and occipital). The Fp1 and Fp2 were excluded from the frontal region averaging for possible eye movement artefacts, even though the subjects had their eyes closed throughout the recording. The signals were visually inspected for any abnormalities/artefacts and were further excluded from the analysis. The basic statistical features, such as maximum, mean, minimum and standard deviation were computed for the above factors. However, there seemed to be a rather large deviation in some instances between the pre-stimulation and post-stimulation for both the control and exposure conditions. As a result, a ratio = post-stim/pre-stim, was therefore


867

(b) Fig. 2. (continued)

computed and the data was represented in Figs. 2 and 3. The ratio’s minimum value described a decrease from before to after exposure and for a maximum value, an increase. 4.2. ECG and PPG results The analysis was conducted by observing the maximum and standard statistical features of ECG and PPG signals. From Figs. 2a and 2d, it can be observed that for ECG recordings throughout the D1 –D4 wavelet subbands, there seemed to be a slight evident increase from before to after exposure at day 3 in comparison to control condition. Also, an evident difference between control and exposure conditions was shown at A4 subband. In relation to PPG analysis in Figs. 2a and 2d, similar increase in coefficient from before to after exposure (exposure condition) was observed at day 3 for all the subbands (D2 –A4 ) except for D1 (standard deviation feature) shown in Fig. 2d. Interestingly, it was shown that an increase (positive ratio) was mainly present for the exposure and a decrease (negative ratio) for the control condition. At day 5, there was a clear difference between exposure and control conditions.

868


(c) Fig. 2. (continued)

For the mean statistical feature of ECG recordings shown in Fig. 2b, there was a complete difference between the exposure and control conditions: throughout all 5 days at D4 subband; difference in day 1, 3, and 5 (D1 , D3 , and A4 subbands); and day 2 and 5 (D2 ). For the PPG recording, the only differences were evident at: day 2 and 3 (D1 ); day 1, 3, and 5 (D2 ); day 2, 3, 4, 5 (D3 ); and day 2 (D4 ), as shown in Fig. 2b (bottom row). For the minimum statistical feature of ECG recordings shown in Fig. 2c, no evident changes were observed, except at A4 subband (day 1 and 3). The PPG recordings showed mainly differences between the increase in exposure and decrease in control for day 3 and 5 (D2 , D3 , D4 , and A4 ). The other distinct changes were shown at D1 (day 2, 3, and 4), as shown in Fig. 2c. 4.3. EEG results The second analysis was performed on the EEG signals. The results extracted from maximum and minimum statistical features (Figs. 3a and 3c) showed a ‘sharp’ increase (positive ratio) in wavelet coefficient for exposure


869

(d) Fig. 2. (continued)

condition at day 3 at frontal and central brain regions (D1 , D2 , and D3 ), with other changes evident between exposure and control at day 1 and 5 (D1 , D2 , and D3 ) at frontal and central region. The additional brain regions, pariatel and occipital exhibited an increase in exposure at day 3 (D3 and D4 ) and a ‘transition’ between the increase to decrease at exposure condition from day 1 to 4 (D1 and D2 ) at pariatel region, as shown in Figs. 3a and 3c. This increase to decrease exposure transition was also evident for the results of standard deviation features in Fig. 3d. At the D1 and D2 subbands for all brain regions, there seemed to be a clear difference between the exposure (increase) and control conditions from day 1 to 3. However, at day 4, the exposure was decreased on that day and the next day 5, it increased identically as the control condition. The other subbands (D3 , D4 , and A4 ) revealed a repeated increase in exposure at

870


(a) Fig. 3. Extracted (a) maximum, (b) mean, (c) minimum, and (d) standard deviation statistical features of exemplary EEG recordings. The statistical feature of the wavelet coefficients in each subband is representated. The figure’s labels are as follows: x-axis represent the days of recordings; y-axis—ratio (after/before); legend—control/exposure; columns—EEG regions; and rows represent wavelet coefficients in each subband.

all the brain regions. The mean statistical feature results revealed a different picture. There was an increase in control condition at: day 3 at pariatel and occipital region (D4 ) and central (D3 ); and day 2 (D3 ). The other related results from Fig. 3b were irregular to conclude any evident changes. The wavelet decomposed components, A4 are within the delta EEG band (0.5–4 Hz), D4 within theta band (4– 8 Hz), D3 within alpha band (8–13 Hz), and D2 within the beta band (13–30 Hz). The lower level decompositions correspond to higher frequencies have negligible magnitudes in normal EEG. 5. Discussion and conclusion Compared to previous studies which applied various linear and nonlinear analysis, such as PSD and approximate entropy techniques to analyse the ECG and EEG results using conventional HRV and EEG frequency bands


871

(b) Fig. 3. (continued)

[14–25], our study’s analysis has obtained promising results in determining the changes in recorded signals. However, more work is required to answer the question on the relation and correlation of wavelet’s extracted subbands with traditional frequency domain bands of HRV and PPG. Perhaps, a feature extraction using wavelet packets for signal decomposition into standard bands would be a viable solution. The wavelet coefficients were computed in D1 –D4 and A4 subbands, since the wavelet coefficients in the further bands (5th, 6th, . . . , etc.) were very close to zero. Since classification is more accurate when the pattern is simplified through representation by important features, feature extraction and selection play an important role in classifying systems such as neural networks. Feature extraction is the determination of a feature or a feature vector from a pattern vector. Feature selection is an optional stage, whereby the feature vector is reduced in size including only, from the classification viewpoint, what may be

872


(c) Fig. 3. (continued)

considered as the most relevant features required for discrimination. In the present study, feature extraction from the PPG, ECG, EEG signals was performed by usage of the DWT. The computed wavelet coefficients can be used as the representing features of the PPG, ECG, EEG signals. These features can be used as inputs of the different neural network architectures. Further work can be performed for improving the classification accuracies by the usage of the wavelet coefficients and/or different feature extraction methods and neural network architectures. Overall, the experimental results revealed that depending on the day of exposure, the subject might exhibit electrophysiological changes to ELF PEMF exposures. In this pilot study, the wavelet feature extractions of different subbands and analysis revealed that the mid-week (day 3) was the time that a substantial increase in ELF PEMF exposure was most evident in comparison to control ‘no exposure’ condition. Further extension of this study is required with additional subjects to be recruited and tested to determine a possible statistical significance outcome of this ELF PEMF exposure time-window effect on the human body.


873

(d) Fig. 3. (continued)

References [1] H. Kordylewski, D. Graupe, K. Liu, A novel large-memory neural network as an aid in medical diagnosis applications, IEEE Trans. Inform. Technol. Biomed. 5 (3) (2001) 202–209. [2] N. Kwak, C.-H. Choi, Input feature selection for classification problems, IEEE Trans. Neural Networks 13 (1) (2002) 143–159. [3] E.D. Übeyli, ˙I Güler, Feature extraction from Doppler ultrasound signals for automated diagnostic systems, Comp. Biol. Med. 35 (9) (2005) 735–764. [4] S.C. Saxena, V. Kumar, S.T. Hamde, Feature extraction from ECG signals using wavelet transforms for disease diagnostics, Int. J. Syst. Sci. 33 (13) (2002) 1073–1085. [5] S.Y. Foo, G. Stuart, B. Harvey, A. Meyer-Baese, Neural network-based EKG pattern recognition, Eng. Appl. Artif. Intell. 15 (2002) 253–260. [6] I. Daubechies, The wavelet transform, time-frequency localization and signal analysis, IEEE Trans. Inform. Theory 36 (5) (1990) 961–1005. [7] P.S. Addison, J.N. Watson, G.R. Clegg, M. Holzer, F. Sterz, C.E. Robertson, Evaluating arrhythmias in ECG signals using wavelet transforms, IEEE Eng. Med. Biol. 19 (5) (2000) 104–109. [8] S. Soltani, On the use of the wavelet decomposition for time series prediction, Neurocomputing 48 (2002) 267–277. [9] M. Akay, Wavelet applications in medicine, IEEE Spectrum 34 (5) (1997) 50–56. [10] M. Unser, A. Aldroubi, A review of wavelets in biomedical applications, Proc. IEEE 84 (4) (1996) 626–638.

874


[11] H. Adeli, Z. Zhou, N. Dadmehr, Analysis of EEG records in an epileptic patient using wavelet transform, J. Neurosci. Methods 123 (1) (2003) 69–87. [12] N. Hazarika, J.Z. Chen, A.C. Tsoi, A. Sergejew, Classification of EEG signals using the wavelet transform, Signal Process. 59 (1) (1997) 61–72. [13] M. Akay, Time Frequency and Wavelets in Biomedical Signal Processing, Institute of Electrical and Electronics Engineers, New York, 1998. [14] A.A. Marino, R.O. Becker, Biological effects of extremely low frequency electric and magnetic fields: A review, Physiol. Chem. Phys. 9 (2) (1977) 131–147. [15] G.B. Bell, A.A. Marino, A. Chesson, Alterations in brain electrical activity caused by magnetic fields: detecting the detection process, Electroencephalogr. Clin. Neurophysiol. 83 (1991) 389–397. [16] D. Cvetkovic, I. Cosic, Automated ELF magnetic field stimulation of the human EEG activity, Int. Comp. Aided Eng. 13 (4) (2006) 313–328. [17] G.B. Bell, A.A. Marino, A. Chesson, Frequency-specific responses in the human brain caused by electromagnetic fields, J. Neurol. Sci. 123 (1994) 26–32. [18] G.B. Bell, A.A. Marino, A. Chesson, Frequency-specific blocking in the human brain caused by electromagnetic fields, NeuroReport 5 (1994) 510–512. [19] A.A. Marino, G.B. Bell, A. Chesson, Low-level EMFs are transduced like other stimuli, J. Neurol. Sci. 144 (1996) 99–106. [20] K. Hausser, D. Tellschaft, F. Thoss, Influence of an alternating 3 Hz magnetic field with an induction of 0.1 mT on chosen parameters of the human occipital EEG, Neurosci. Lett. 239 (1997) 57–60. [21] E.B. Lyskov, J. Juutilainen, V. Jousmaki, J. Partanen, S. Medvedev, O. Hanninen, Effects of 45 Hz magnetic fields on the functional state of human brain, Bioelectromagnetics 14 (1993) 87–95. [22] S. Ghione, C.D. Seppia, L. Mezzasalma, L. Bonfiglio, Effects of 50 Hz electromagnetic fields on electroencephalographic alpha activity, dental pain threshold and cardiovascular parameters in humans, Neurosci. Lett. 382 (2005) 112–117. [23] A. Sastre, C. Graham, M.R. Cook, M.M. Gerkoch, P. Gailey, Human EEG responses to controlled alterations of the earth’s magnetic field, Clin. Neurophys. 113 (2002) 1382–1390. [24] Z. Tabor, J. Michalski, E. Rokita, Influence of 50 Hz magnetic field on human heart rate variability: Linear and nonlinear analysis, Bioelectromagnetics 25 (2004) 474–480. [25] E. Baldi, C. Baldi, B.J. Lithgow, A pilot investigation of the effects of extremely low frequency pulsed electromagnetic fields on human’s heart rate variability, Bioelectromagnetics 28 (2007) 64–68.

Dean Cvetkovic (Ph.D. 2005 in biomedical engineering, RMIT University). He is a Postdoctoral Research Fellow at School of Electrical and Computer Engineering, RMIT University. His research is based on investigation of extremely low frequency electromagnetic field, photonic and audio stimulation on the human brain wave activity and skin impedance. He has 9 years experience in the research and development areas of biomedical engineering, brain–computer interface, EEG/ECG and sleep monitoring, bioinstrumentation, bioelectromagnetism, biosignal and biostatistical processing and analysis, biomedical artificial intelligence, automation and modeling. His current research projects include the design and development of biofeedback system for human sleep induction and relaxation, automated sleep scoring and monitoring systems in obstructive sleep apnoea/hypopnoea detection, mathematical sleep modeling/simulation and investigation of ELF/RF EMF bioeffects, etc. He has published over 30 refereed publications. Elif Derya Übeyli (Ph.D. 2004 in electronics and computer technology, Gazi University). She is an Associate Professor at the Department of Electrical and Electronics Engineering, TOBB University of Economics and Technology. She has worked on variety of topics including biomedical signal processing, neural networks, optimization and artificial intelligence. She has worked on several projects related with biomedical signal acquisition, processing and classification. Dr. Übeyli has served (or is currently serving) as a program organizing committee member of the national and international conferences. She is editorial board member of several scientific journals. She is assistant editor of Expert Systems. She is serving as a guest editor to the Expert Systems on a special issue “Advances in Medical Decision Support Systems.” Moreover, she is voluntarily serving as a technical publication reviewer for many respected scientific journals and conferences. She has also published 78 journal and 33 conference papers on her research areas. Irena Cosic (Ph.D. 1985 in biomedical engineering, University of Belgrade). From 1977 to 1989 she was senior researcher at the Institute Vinca, Belgrade, where she was working on a variety of projects in biomedical engineering, digital signal analysis and telecommunication. In 1980 she commenced her research in digital signal processing applications on linear macromolecules. This research resulted in the Resonant Recognition Model (RRM) of protein and DNA interactions. From 1989 until the beginning of 1993 she was a Research Fellow in the Biochemistry Department, Monash University, where she was able to test some practical applications of the RRM model. From 1993 until February 2002 she has been a Senior Lecturer/Associate Professor in the Department of E&CS at Monash University, where she is continuing her research in the field of biomolecular electronics. Since February 2002 she is appointed as Professor of Biomedical Engineering and Head of School of Electrical and Computer Engineering at RMIT University. Professor Cosic is senior member of IEEE, Fellow of IEAust and active member of a number of other national and international professional societies. She teaches bioelectromagnetism and biomedical engineering, she has published one research book, one international patent as well as over 100 other refereed publications predominantly in the area of biomolecular electronics and biomedical engineering.