Neural network for photoplethysmographic respiratory rate monitoring

# Bio-Optics in Medicine

Neural network for photoplethysmographic respiratory rate monitoring A. Johansson

Department of Biomedical Engineering and Swedish National Centre of Excellence for Non-invasive Medical Measurements (NIMED), Link6pings Universitet, Link6ping, Sweden

Abstract--The reflection mode photoplethysmographic (PPG) signal was studied with the aim of determining respiratory rate. The PPG signal includes respiratory synchronous components, seen as frequency modulation of the heart rate (respiratory sinus arrhythmia), amplitude modulation of the cardiac pulse and respiratoryinduced intensity variations (RIIVs) in the PPG baseline. PPG signals were recorded from the foreheads of 15 healthy subjects. From these signals, the systolic waveform, diastolic waveform, respiratory sinus arrhythmia, pulse amplitude and RIIVs were extracted. Using basic algorithms, the rates of false positive and false negative detection of breaths were calculated separately for each of the five components. Furthermore, a neural network was assessed in a combined pattern recognition approach. The error rates (sum of false positive and false negative breath detections) for the basic algorithms ranged from 9.7% (pulse amplitude) to 14.5% (systolic waveform). The corresponding values for the neural network analysis were 9.5-9.6%. These results suggest the use of a combined PPG system for simultaneous monitoring of respiratory rate and arterial oxygen saturation (pulse oximetry). Keywords--Photoplethysmography, Optical sensors, Pulse oximetry, Respiratory rate, Ventilation monitoring, Neural networks

\

Med. Biol. Eng. Comput., 2003, 41,242-248

1 Introduction

THE PHOTOPLETHYSMOGRAPHIC(PPG) signal is obtained by measuring the intensity of light penetrating through or reflected by the skin. The dynamic part of the signal is caused by variations in blood volume and perfusion of the tissue affecting the scattering and absorption of the incident light (CHALLONER, 1979). The non-invasive character and the minimum disturbance of the variables under observation have made the technique widely used. The most widespread application of PPG is in the monitoring of arterial oxygen saturation by pulse oximetry (AOYAGI et al., 1974; TREMPERand BARKER,1989), where the cardiac synchronous component of the PPG signal is used. Other applications of the cardiac-related component include cutaneous blood flow measurement and heart rate monitoring (CHALLONER, 1979; ELDRUP-JORGENSEN et al., 1966; LINDBERGet al., 1992). A respiratory-induced intensity variation (RIIV) of the PPG signal baseline is present (LINDBERGe t al., 1992). The RIIV is not fully understood, but is believed to be caused by skin blood volume fluctuations induced by the respiratory variations in intra-thoracic pressure transmitted to the measurement

Correspondence should be addressed to Dr Anders Johansson; emaih [email protected] Paper received 9 July 2002 and in final form 11 November 2002 MBEC online number: 20033747 © IFMBE: 2003 242

J site by the venous system (DORLAS and NIJBOER, 1985; UGNELL, 1996; JOHANSSONand 0BERG, 1999b). By extracting the RIIV and the cardiac synchronous component, a photoplethysmographic system for monitoring heart and respiratory rates (LINDBERGet al., 1992) and tidal volumes (JOHANSSON and 0BERG, 1999a) has been presented and evaluated clinically (NILSSON et al., 2000). Respiratory synchronous variations originating from the arterial side of the circulatory system are also present in the PPG signal. These include the well-known respiratory sinus arrhythmia (RSA), primarily a vagal heart rate reflex, and the amplitude modulation of the cardiac pulse, caused by cardiac stroke volume variations as a result of respiratory venous return variations and blood pooling in the pulmonary circulation (KAMATHand FALLEN,1993; DORNHORSTet al., 1952; BUDA et al., 1979). The respiratory rate would be more accurately derived from the PPG signal if the arterial respiratory information was utilised in addition to the RIIV. in a neural network (NN), a structure of mathematical units (neurons) is constructed. Each neuron calculates the sum of its inputs and uses an activation function for its output. The neurons are connected by weights that are adjusted in a training process, aiming at minimising the network error function, in the training process, the backward propagation algorithm is often used (BISHOP, 1995). The main advantage of a neural network is the possibility to generate a complex decision boundary without full knowledge of the statistical properties of the data under observation. This has made the method attractive in biomedical applications, and NNs are used in fields such as the diagnosis of Medical & Biological Engineering & Computing 2003, Vol. 41

infrared light emitting diodes (940 nm) •

~

,'/

ght barriers 10 mm I ,hotodetectors 'PG reflectance sensor low reference (Optovent)

icone backing

Fig. 1 Measurement set-up presenting photoplethysmographic reflection mode sensor emitting infrared light of 940 nm and respiratory ai~tow reference (Optovent)

vascular disease from pulse waveforms (ALLEN and MURRAY, 1993) and ECG classification (WATROUS and TOWELL, 1995). in ventilation monitoring, NNs are used to detect breathing circuit problems (SITTIG and ORR, 1992), in ventilation mode recognition (LEONand LORINI, 1997) and to classify effective or ineffective respiration in neonates (WILKS and ENGLISH, 1995). The aim of this study is to evaluate the accuracy in respiratory rate measurements of five different respiratory synchronous components of the PPG signal, and to see whether a neural network, combining the components, can further improve the accuracy in the detection of breaths.

2 Methods 2.1 Subjects and measurement set-up

Fifteen healthy subjects (11 men and four women, aged 23-42 years) participated in the study. The measurements were taken with the subjects in the supine position. A PPG reflection mode sensor was constructed (Fig. 1) and positioned laterally on the forehead using a non-elastic Velcro head band, asserting negligible pressure on the sensor. The sensor included eight light emitting diodes emitting infrared light of 940 nm and two photodetectors. The optical components were mounted in medical-grade silicone. Light barriers were included to avoid direct transmission of light between light sources and photodetectors. The detector-diode distances were 3.5 and 5 mm, respectively. The detected signal was offset balanced and amplified (LrNDBERGet al., 1992). As a respiratory reference, airflow humidity was sensed* (VEGFORS et al., 1994). The sensor detects respiratory variations in the humidity accumulating on an optical fibre end positioned in front of the airways. The reference signal was delivered from the device as a square wave. The PPG and the reference signal were AD-convertedt (100 Hz) and stored on a PC.

consecutive sounds. The subjects were told to follow the pace set by the metronome.

2.3 Feature extraction Three 2min PPG periods, free from movement-induced disturbances, were selected from each measurement, resulting in 12 min of signals from each subject. The RIIV component of each 2 min period was extracted using a 16th-order bandpass Bessel filter (0.13-0.48Hz (UGNELL, 1996)), and the cardiac synchronous component was extracted using a 5th-order bandpass Butterworth filter (0.50-2.0Hz). From the extreme values of the cardiac signal obtained, each peak and valley was detected, and, from the peaks i of each 2 min period, the following components were formed (Fig. 2): C t : S Y S T i = peak value (systolic waveform)

(1)

C 2 : D I A S T i = previous valley value (diastolic waveform) (2) C 3 : R S A i = time since last peak t (respiratory sinus

arrhythmia) C4:PULSE i = SYST i -DIAST

(3)

i (pulse amplitude)

(4)

Furthermore, a discrete version of the RIIV was designated as Cs:RIIV~ = R I I V a t time of SYST~

(5)

PPG baseline (RIIV)

~YST~ PPG cardiac synchronous signal

2.2 Measurements After a resting period of 5-10 min, two measurements were taken from each of the subjects (i) 10min of spontaneous breathing without instructions (ii) 10 min of metronome-controlled breathing. For the second measurement, a computerised metronome sounding at a randomised rate was constructed. The metronome rate was uniformly distributed within the interval 6-18min -1, but no larger shifts than 2min -1 were allowed between two *Optovent system tDAQCard-1200, National Instruments, USA

Medical & Biological Engineering & Computing 2003, Vol. 41

J DIAS7

time, s

1 2 3 0 Fig. 2 Extraction of respiratory synchronous components o f photo-

plethysmographic signal Five values from each cardiac pulse (denoted i) are extracted." systolic value SYSTi, diastolic value DIAST,~ respiratory sinus arrhythmia RSAi, pulse amplitude PULSEi, and RIIV value RII~. 243

Xl

1,1,1

SYST

1•

•

DIAST

B1,1

.

0 •

•

W1.1

/31,2

Y1 (1 for inspiration)

•

RSA

• •

0 •

:~2,1

W1,2s,1 B1,3

Y2 (1 for expiration)

•

PULSE

B2,2 0

w2,s,2

• RflV

1

•"

•

W1,2S,S 0 ii+l i+2i+'3i+4

I BI,s

x2s

Fig. 3 Fully connected feed-Jbrward neural network with 25 inputs', 2 outputs" and 5 neurons in hidden laye~ Activation functions of sigmoid type and bias values are used throughout. At specific time i and event (inspiration, expiration, or neither of two), five consecutive values of each component were normalised and used as input to network

2.4 Separate analysis Analyses of the five respiratory components Cj (j = 1 , . . . , 5) were performed separately. The baseline trend of each 2 min period was removed by subtracting the smoothed function of the signal, yielding C;, i according to CJ'i = C j ' i -

Cj, i 4 + " " + Cj, i+4 9

(6)

Each period was also normalised to m e a n = 0 and SD = 1 (yielding C~i). Two simple algorithms were then applied for breath detection (a) zero-crossing algorithm: inspiration/expiration when

d~ii>0 and C~,i t < 0

(7)

and the opposite for expiration/inspiration (b) p e a k algorithm: inspiration/expiration when djli>Cjli

t and C ; ' , i > C ; i i + t

(8)

and the opposite for expiration/inspiration. Whether a true expression represents inspiration or expiration depends on the time-course of the specific component. From the reference, each inspiration and expiration was noted, and the false positive and false negative breath detection rates were calculated for the five components and for both algorithms. 2.5 Combined analysis: neural network topology and training

The components C1-C5 were presented to a neural network as the values from five consecutive heartbeats after a specific time i. 244

The five values of each component were normalised to a range of 0-1. An output of [1 0] was assigned when the inputs followed inspiration, [0 1] was assigned when the inputs followed expiration, and [0 0] was assigned when the input, followed neither of the two. Shifting the five signals one heartbeat, performing anew normalisation and assigning a new output formed the next pattern, in this way, typically, 60 patterns per minute of measurement time were obtained. Consequently, a network with 25 inputs and two outputs was formed. A network structure with one hidden layer consisting of five neurons was chosen, resulting in a 25 : 5 : 2 topology (Fig. 3). The network was a fully connected feed-forward network, with sigmoid activation functions and bias values used throughout. All neural network processing was performed in the MATLAB environmentS. For each subject (12min), the network was trained with the signals of the remaining subjects (14xl2min). Backpropagation was used in the minimisation of the error function, and the training was repeated for 100 epochs. To avoid local minima, the training was restarted if the sum-squared error was above a specific value (equal for all 15 training sets).

2.6 Combined analysis: neural network evaluation After the training procedure, the signals of the subject not involved in the training were presented to the network, and the output from the network was compared with the reference. Only the first output of the network (representing inspiration) was used (see Section 4). Breaths were assigned at peaks above an appropriate threshold level. ~The M a t h W o r k s

Medical & Biological Engineering & Computing 2003, Vol. 41

cardiac signal

3.2 Neural network analysis

SYST (01) DIAST (02) RSA (Cs) PULSE

(C4)

RllV (c O

80s 78

Fig. 4

reference signal • = insp 120s 117

time, s heartbeat number

Characteristic example of photoplethysmographic cardiac synchronous signal and reference signal (time as x-scale) during metronome-controlled breathing. Inspirations are marked. Five calculated respiratory components" are included (heartbeat number as x-scale); y-axes are in arbitrary units"

The total number of patterns was 10329, with 2041 following inspiration, 2029 following expiration, and 6259 following neither of the two. An overview of the patterns is presented in Fig. 5. The general effects of an inspiration were a decrease in S Y S T and an increase in DIAST. This caused the PULSE parameter to fall. RSA increased after an inspiration, and RIIV decreased. The opposite was seen at expiration. The network weights after training are summarised in Table 2. The numbers presented are absolute mean values over the 15 training sessions and over the five neurons in the hidden layer, it can be noticed that the weights ofDIAST, RSA and PULSE are the highest, corresponding well with the separate analysis, and that the weights are highest for input values corresponding to heartbeats 1-3 after the respiratory event. The neural network evaluation results are presented in Table 3, and a typical network output is shown in Fig. 6. In the Table, the numbers of false positive (FP) and false negative (FN) breath detections are presented for both measurements, together with the total number of reference breaths n and the threshold tr used in the detection process. The error rate was similar for both spontaneous and metronome-controlled breathing, with an error occurring at the rate of 9.5-9.6% of the breaths. This is slightly better than, but in the same range as, the PULSE parameter when used alone.

3 Results The mean heart rate was fotmd to be 60.4 (-4-9.4 SD t min -1 (n = 15). The mean respiratory rate was 11.0 (3.6) m i n - for the measurements of spontaneous breathing and 11.6 (0.9) min -1 for the metronome measurements. A characteristic example of the five respiratory components is presented, together with the reference, in Fig. 4.

3.1 Separate analysis

The results of the separate analysis of the five components for both algorithms are presented in Table 1. The columns show the mean (SD) rates of falsely detected breaths. No significant difference in the error rates between the measurements of spontaneous breathing and metronome controlled breathing was found. When the error rates (the sum of false negatives and false positives) were studied, the zero-crossing algorithm performed better than the peak algorithm for all components. The best results were fotmd for the pulse amplitude, with an error rate of 9.9% (5.2% false positives and 4.7% false negatives). The least accurate component was the systolic waveform, with an error rate of 15%.

1

1

Fig. 5

2 3 4 5 inspiration n=2041

1

2 3 4 5 expiration n=2029

1

2

3 4 other n=6259

5

Patterns for neural-network analysis. Boxes" show lower quartile, median and upper quartile values'. Patterns are divided according to output (inspiration, expiration and others'); x-scales are heartbeat numbers" after event

Table 1 Results" of separate analysis of five PPG respiratory synchronous components'. Zero-crossing algorithm and peak algorithm for breath detection are used. Columns show mean (SD) rates of falsely detected breaths (n -- 15). Results of spontaneous and metronome-controlled breathing are put togethe~ and lowest error rate is in bold type

Zero-crossing algorithm False positives, % SYST DIAST

RSA PULSE RIIV

11.8 8.5 3.7 5.2 7.7

Peak algorithm

False negatives, %

(0.7) (0.7) (0.4) (0.5) (0.5)

3.2 3.0 6.9 4.7 5.9

(0.3) (0.4) (0.5) (0.4) (0.4)

False positives, % 30.3 22.9 12.5 14.7 18.5

(1.7) (1.9) (1.1) (1.3) (1.0)

False negatives, % 0.4 0.8 2.1 1.1 1.9

(0.04) (0.1) (0.2) (0.1) (0.1)

Total number of reference breaths - - 2136 Medical & Biological Engineering & Computing 2003, Vol. 41

245

Table 2 Summary o f weights o f trained networks'. Numbers" presented are absolute mean values of weights over 15 training sessions and over five neurons in hidden layer

Heartbeat number

SYST

DIAST

RSA

PULSE

RIIV

Mean

1 2 3 4 5 Mean

0.80 0.97 0.82 0.80 0.95 0.87

1.38 1.4 0.96 1.06 0.85 1.13

1.17 1.25 1.11 1.05 0.95 1.11

1.07 1.07 1.26 0.87 1.07 1.07

0.87 0.75 0.88 0.94 0.9 0.87

1.06 1.09 1.01 0.94 0.94

Table 3 Results" o f neural network analysis combining five PPG respiratory synchronous components'. Number o f fi~lse positive (FP) and false negative (FN) breath detections are presented for both measurements', together with total number of reference breaths (n) and threshold (to used in detection process

Spontmaeous breathing Subject

n

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Total

~._o~

1

93 71 55 72 46 57 62 106 71 76 58 97 64 29 36 993

~

Metronome-controlled breathing

FP

FN

1 0 3 4 4 3 3 1 5 1 5 1 5 10 16 62 6.2%

1 0 0 4 0 0 9 1 6 0 0 7 0 5 1 34 3.4%

reference

~_ 0

1

heartbeat number

114

(0)

(time, s)

(120)

Typical 2 min neural-network output (Ysnspiration). Peaks" Fig.6 above (- - -) threshold are regarded as inspirations. Reference (inspiration) points" are marked," x-scale is heartbeat number

4 Discussion There are two main findings of this study: first, the P U L S E parameter is the most reliable component when respiratory rate is determined from the PPG signal, with mistakes (false positives or false negatives) in 9.9% of the breaths, as detected by the reference: and, secondly, the different respiratory components o f the PPG signal can be combined by pattern recognition for improved accuracy. An advantage o f applying the P U L S E component, not addressed in this study, is that it originates in the 'high-pressure' arterial system and is probably less affected by motion (hydrostatic) disturbance (JOHANSSONet al., 1999). A system based on this component would therefore be preferable to one based solely on the RIIV signal, which originates in the 'low-pressure' venous system. However, the RIIV signal should be included as 246

tr

n

0.05 0.10 0.10 0.10 0.15 0.18 0.40 0.05 0.50 0.10 0.10 0.10 0.40 0.40 0.50

84 65 64 70 69 66 67 69 63 73 67 77 73 69 72 1048

FP

FN

1 4 0 6 1 2 6 5 8 1 4 4 2 2 0 46 4.4%

3 1 0 9 1 0 2 4 6 1 1 0 4 6 15 53 5.1%

an input signal in a pattern recognition system, as it is largely different in origin from the other components. The neural network suggested in this study only makes slightly fewer errors (9.5-9.6%) than the P U L S E parameter acting alone. The physiological origins of the components in use are not very well known, and the error rate could perhaps be reduced by removing the most dependent and inaccurate components, by increasing the number o f patterns (helping the training) or by implementing a signal selection process (making training not necessary). More advanced filters and an optimised network structure are other important sources for improvement. A modified Bessel filter for extracting the PPG cardiac synchronous part has been presented (UGNELL, 1995), and promising results have been found when wavelet RIIV extraction is applied (unpublished results). Furthermore, a comparison between the separate and the combined analysis is not completely straightforward, as the separate analysis has the advantage o f working with complete periods (in the smoothing and normalisation), whereas the network analysis is performed in real-time. The forehead was chosen as the measurement site as (i) the photoplethysmograms from the forehead accurately reflect arterial and venous blood volume variations, as nervous effects are minimal (ZIEGE et al., 1997) (ii) in reflectance pulse oximetry, the forehead is an appropriate choice (MENDELSON, 1992) (iii) photoplethysmograms o f high amplitude and signal-tonoise ratio are obtained at this location, owing to high Medical & Biological Engineering & Computing 2003, Vol. 41

vascularisation and the frontal bone acting as a reflector (TUN et al., 1983; NIJBOER and DORLAS, 1982) (iv) the sensitivity to patient motion is reduced at this site. For these reasons, a more traditional transmission measurement approach was not considered. The absorption and scattering of skin and blood are at a minimum in the infrared region. Thus, by using infrared light sources, a larger sample volume is seen, and a more reliable PPG signal is obtained owing to the averaging effect and a larger fraction of light reaching vessels localised deeper in the skin. Measurements were performed during spontaneous and metronome-controlled breathing. The latter was used to produce different breath-to-breath intervals, and thus a more variable material. Neither in the separate nor in the combined analysis was a significant difference between the error rates of the two measurement types noticed. The Optovent system was chosen as the reference method because it measures a variable related to the respiratory gas flow without the need of a mouthpiece (e.g. pneumotachographs or spirometers are known to alter the respiratory pattern (GILBERT et al., 1972)). Chest expansion might be even more closely related to the PPG respiratory variations than the respiratory gas flow. indirect monitoring of respiration, e.g. transthoracic impedance plethysmography, was, however, not considered reliable enough as reference in this study (JOHANSSON et al., 1999; NILSSON et al., 2000). The selection of six 2 min periods for each subject was straightforward, as the motion disturbances were easily identified (visually). The first periods free from this type of disturbance were selected. As all five respiratory components were extracted from the same PPG signal, motion disturbance was of minor interest in this study, as it was the relationship between the components that was under investigation. A disturbance in the PPG signal therefore affected the components equally. Furthermore, if non-relevant data were included in the neural network training, a convergence of the network error function would have been difficult to achieve. When we were attempting to combine the respiratory components, pattern recognition in the form of a neural network was chosen, as the relationship between the different respiratory components of the PPG signal is complex. The decision boundary could thus be formed by the network. The implementation in a monitoring system is straightforward, and the speed of processing (once the network is trained) is high. The number of inputs were limited to five for each component of the PPG signal, as five heartbeats represent a reasonable amount of time for the effects of each event to be captured fully. This assumption was verified when the weights of the trained networks were studied (Table 2). One hidden layer of neurons was assumed sufficient for this problem. This layer included five neurons, chosen from the general rule that the number of hidden neurons should be the 2-logarithm of the number of inputs (WILKS and ENGLISH, 1995). The training was stopped after a fixed number of epochs and restarted if the error was above a specific level. This procedure was used to avoid local minima and to avoid over-fitting the network to the training set. By shifting the signals by one heartbeat, a new pattern was formed. This caused each value to be seen by the network five times. However, as a new pattern was generated after each shift, the effects of redundancies are assumed to have a minor influence on the results. The PULSE parameter inputs were actually a linear combination of the inputs from SYSTand DIAST. This parameter was included anyway, as it was plausible that this form of pre-processing would make it easier for the network to converge. This was confirmed when this parameter made the least number of errors in the separate analysis. The patterns following expiration were less clear than those following inspiration. Therefore only the first output of the Medical & Biological Engineering & Computing 2003, Vol. 41

network, representing inspiration, was used in the network evaluation, it may be possible further to improve the system by combining the two outputs. Another important part was the determination of a suitable threshold to interpret the network output, in the present evaluation, this threshold had to be varied, as the networks were different for each subject. The subjects in this study were young and healthy. How a pattern recognition approach would work in a pathological condition, or in elderly patients, remains to be investigated. Our belief is that, in situations where some components are weak or disturbed (for instance, a less pronounced RSA in an elderly patient), the network will recognise the respiration from the other components. To do this effectively, it is of great importance that the neural network should have been trained by relevant data recorded in the specific environment. Two interesting furore aspects of the technique are to develop selftraining patient-specific networks and to include the amplitudes of the respiratory components to estimate tidal volume.

5 Conclusions Photoplethysmography is an interesting approach for ventilation monitoring, as this technique would make simultaneous monitoring of respiratory rate and arterial oxygen saturation (pulse oximetry) possible, minimising the number of sensors attached to the patient. When using photoplethysmography for respiratory rate monitoring, an accurate algorithm for the detection of breaths is demanded. The photoplethysmographic signal contains several respiratory synchronous components of different origins. The pulse amplitude of the cardiac component (representing systolic-to-diastolic variation) seems to be the most reliable parameter for breath detection. The accuracy may be further improved by combining the information from the different respiratory components in a pattern recognition system, e.g. a neural network implementation.

References

ALLEN,J., and MURRAY,A. (1993): 'Development of a neural network screening aid for diagnosing lower limb peripheral vascular disease from photoelectric plethysmography pulse waveforms', Physiol. Meas., 14, pp. 13-22 AOYAGI,T., KISHI, M., YAMAGUCHI,K., and WATANABLE,S. (1974): 'Improvement of the earpiece oximeter', Proc. 13th Ann. Meet. Japanese Soc. Med. Electron. Biol. Eng.

BISHOP, C. M. (1995): 'Neural networks for pattern recognition' (Oxford University Press, Oxford, 1995) BUDA, A. J., PINSKY, M. R., INGELS, N. B., DAUGHTERS, G. T., STINSON, E. B., and ALDERMAN,E. L. (1979): 'Effect of intrathoracic pressure on left ventriculax performance', N. Engl. J. Med., 301, pp. 453-459 CHALLONER, A. V J. (1979): 'Photoelectric plethysmography for estimating cutaneous blood flow' in ROLFE,P. (Ed.): 'Non-invasive physiological measurements' (Academic Press, London, 1979), pp. 125-151 DORLAS, J. C., and NIJBOER,J. A. (1985): 'Photo-electric plethysmography as a monitoring device in anaesthesia', B~ J Anaesth., 57, pp. 524-530 DORNHORST, A. C., HOWARD, R, and LEATHART, G. L. (1952): 'Respiratory vaxiations in blood pressure', Cireulation, 6, pp. 553-558 ELDRUP-JORGENSEN,S., SCHWARTZ,A. I., and WALLACE,J. D. (1966): 'A method for clinical evaluation of peripheral circulation: photoelectric hemodensitometry', Surgery, 59, pp. 505-513 GILBERT, R., AUCHINCLOSS, J. H., BRODSKY, J., and BODEN, W. (1972): 'Changes in tidal volume, frequency, and ventilation induced by their measurement', J. Appl. Physiol., 33, pp. 252-254

247

JOHANSSON, A., and ~)BERG,P. A. (1999a): 'Estimation of respiratory volumes from the photoplethysmographic signal--Paxt I: experimental results', Ned. Biol. Eng. Comput., 37, pp. 42-47 JOHANSSON,A., and OBERG, R A. (1999b): 'Estimation of respiratoryvolumes from the photoplethysmographic signal-Paxt II: a model study', Ned. Biol. Eng. Comput., 37, pp. 48-53 JOHANSSON, A., OBERG, R A., and SEDIN, G. (1999): 'Monitoring of heart and respiratory rates in newborn infants using a new photoplethysmographic technique', J. Clin. Nonit., 15, pp. 461-467 KAMATH,M. V, and FALLEN,E. L. (1993): 'Power spectral analysis of heart rate variability: a noninvasive signature of cardiac autonomic function', Crit. Rev. Biomed. Eng., 21, pp. 245-311 LEON, M. A., and LORINI, E L. (1997): 'Ventilation mode recognition using artificial neural networks', Comput. Biomed. Res., 3D, pp. 373-378 LINDBERG, L.-G., UGNELL, H., and OBERG, R A. (1992): 'Monitoring of respiratory and heart rates using a fibre-optic sensor', Ned. Biol. Eng. Comput., 3D, pp. 533-537 MENDELSON, Y. (1992): 'Pulse oximetry: Theory and applications for noninvasive monitoring', Clin. Chem., 38, pp. 1601-1607 NIJBOER, J. A., and DORLAS, J. C. (1982): 'The origin of inverted waveforms in the reflection plethysmogram', B~ J. Anaesth., 54, pp. 1289-1293 NILSSON, L., JOHANSSON,A., and KALMAN,S. (2000): 'Monitoring of respiratory rates in postoperative care using a new photoplethysmographic technique', J. Clin. Nonit. Comput., 16, pp. 309-315 SITTIG,D. E, and ORe, J. A. (1992): 'A parallel implementation of the backwaxd error propagation neural network training algorithm: experiments in event identification', Comput. Biomed. Res., 25, pp. 547-561 TREMPER, K. K., and BARKER, S. J. (1989): 'Pulse oximetry', Anesthesiology, 70, pp. 98-108

248

TUN, E., TUN, M., MAIBACH, H. I., and GuY, R. H. (1983): 'Basal perfusion of the cutaneous microcirculation: Measurements as a function of anatomic position', J. Investigat. Dermatol., 81, pp. 442-446 UGNELL, H. (1996): 'The respiratory synchronous photoplethysmographic signal. Its dependence on light wavelength and sample volume', Ned. Biol. Eng. Comput., 34, pp. 275-276 VEGFORS, M., LINDBERG, L.-G., PETTERSSON, H., and 0BERG, R A. (1994): 'Presentation and evaluation of a new optical sensor for respiratory rate monitoring', Int. J. Clin. Non. Comput., 11, pp. 151-156 WATROUS, e., and TOWELL,G. (1995): 'A patient-adaptive neural network ECG patient monitoring algorithm', in 'Computers in cardiology' (IEEE, Cat. No. 95CH35874, New York, 1995), pp. 229-232 WILKS, R A. D., and ENGLISH, M. J. A. (1995): 'A system for rapid identification of respiratory abnormalities using a neural network', Ned. Eng. Phys., 17, pp. 551-555 ZIEGE, S., SCHMID-SCHONBEIN, H., GREBE, R., and MARTIN, E. (1997): 'Long-term registration of cutaneous microcirculation during general anaesthesia', Int. J. Nicrocire., 17, pp. 385-394

Author's biography ANDERS JOHANSSONreceived the MSc degree in Applied Physics and Electrical Engineering, in 1996, and the PhD in Biomedical Engineering, in 2000, from LinkGpings Universitet, Sweden. He is currently working as am Assistant Professor in the Depaxtment of Biomedical Engineering at LinkGpings Universitet. His main research interests axe optical sensors for non-invasive measurements in respiration monitoring, otitis diagnosis and cartilage quality assessment, with a special interest in photoplethysmography and spectroscopy.

Medical & Biological Engineering & Computing 2003, Vol. 41