TOWARDS UNOBTRUSIVE AUTOMATED SLEEP STAGE

CLASSIFICATION

Polysomnography using Electrodes on the Face

Igor J. Berezhnoy, Gert-Jan de Vries, Tim Weysen, Jonce Dimov and Gary Garcia-Molina

Philips Research Laboratories, High Tech Campus 34, 5656 AE Eindhoven, The Netherlands

Keywords:

Sleep, Automated Scoring, Comfortable Polysomnography.

Abstract:

Although sleep stage annotation (SSA) is historically known from clinical practice and typically performed

by a certiﬁed expert on the basis of visual examination of polysomnography (PSG) signals. Automatic SSA

has emerged as a tool to assist sleep experts and to accelerate the analysis of PSG data. New advances in

signal processing and sensor technology start to enable the application of SSA in home solutions as well.

In today’s busy lives, sleep plays a central role and good quality sleep helps us to deal with the stress of

everyday life. Being able to enhance sleep quality thus is a major opportunity to help people in reducing the

inﬂuence of stress on their live, health and wellbeing. The advent of consumer products aimed at enhancing

the sleep experience has propelled the need for home sleep monitoring and inducing solutions which can i)

provide automatic SSA using sensors that interfere minimally with the sleep process and ii) provide sleep

stage information in real-time in order to be suitable for closed-loop sleep inducing solutions. In this paper,

we examine two possible alternatives for unobtrusive sleep monitoring. The ﬁrst one uses respiratory, cardiac

and wrist actigraphy signals while the second one relies on Facial PSG electrodes positioned on the facial area

which allow for unobtrusive and comfortable sensors arrangements.

1 INTRODUCTION

Sleep stage scoring in clinical practice results from

manual visual scoring on the basis of Polysomnogra-

phy (PSG) data. Typical PSG setups require afﬁxing

electrodes at various locations of the patient’s body to

capture a number of physiological signals including

electroencephalogram (EEG), cardiac activity, respi-

ratory effort, ocular activity (EOG), and chin myo-

graphic activity (EMG). Consumer applications rely-

ing on sleep monitoring do not require the same level

of accuracy as clinical practice. In addition, the level

of obtrusiveness of traditional clinical PSG setups is

unacceptable in the consumer domain.

In this paper, we consider two minimally obtru-

sive sleep monitoring approaches. The ﬁrst approach,

which we refer to as RHA, is based on the monitoring

of respiratory effort, cardiac signals, and wrist actig-

raphy. The second approach, which we refer to as

Facial PSG, is based on the measurement of signals

from two electrodes positioned on a single side of the

subjects’s face.

To establish a reference for our study, in Table 1

we present average agreement scores between two

sleep stage scoring approaches i) clinical sleep scor-

ing (performed by trained human experts and scoring

according to criteria of (Rechtschaffen and A.Kales,

1968), and ii) an automated scoring techniques. Both

of their approaches use data of the full PSG setup

(EEG, EOG and EMG).

Table 1 shows the level of agreement between

raters in terms of i) Cohen’s Kappa statistics (sec-

ond column) and ii) percentage of agreement (third

column). It is important to note that the reported

values in Table 1 were not obtained from the same

dataset. They originate from various studies: i) av-

erage agreement between two independent sleep ex-

perts (row two) observing the same PSG data has been

reported in (Virkkala, 2005), ii) comparison of au-

tomated techniques has been performed on the data

recorded in this study.

Table 1: Sleep stages classiﬁcation agreement ﬁgures.

Cohen’s Kappa Agreement(%)

Human expert(PSG) vs. Hu-

man expert (PSG)

0.80 86.00

Automated technique A

(PSG) vs. Automated

technique B (PSG)

0.69 85.48

487

J. Berezhnoy I., de Vries G., Weysen T., Dimov J. and Garcia-Molina G..

TOWARDS UNOBTRUSIVE AUTOMATED SLEEP STAGE CLASSIFICATION - Polysomnography using Electrodes on the Face.

DOI: 10.5220/0003890104870492

In Proceedings of the International Conference on Health Informatics (BSSS-2012), pages 487-492

ISBN: 978-989-8425-88-1

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

From the agreement ﬁgures shown in the Ta-

ble 1 we can conclude that agreement between human

raters is signiﬁcantly higher than that of automated

techniques with expert annotation despite operating

on the same kind of input data (PSG).

In this paper we compare performances of the

RHA and Facial PSG approaches on the task of auto-

mated epoch bases sleep state annotation. This paper

is organized as follows. Section 2 describes the exper-

imental setup. In Section 3 we present the signal pro-

cessing and machine learning methods that were used

in this study. The results are discussed in Section 4.

Finally Section 5 summarizes the main conclusions of

this work and proposes future research directions.

2 EXPERIMENT DESCRIPTION

This study considers two data sets: First and Second,

acquired in two separate studies. The First dataset in-

cludes respiratory effort, cardiac, and acti-graphy sig-

nals. The RHA approach was tested on the First data

set.

The Second dataset was used for validation of the

Facial PSG approach which had as a major objective

to show that sleep state estimation can be performed

with electrodes applied on a single side of the face

(for which we used EOG-Left placed above the eye

and mastoid reference). Further details on the datasets

are provided in the next sections.

2.1 FirstDataset

The First data set contains overnight PSG recordings

of six young healthy volunteers (mean age 27 y.o., 4

males and 2 females). In a screening phase, selec-

tion of participants was based on absence of subjec-

tive sleep complaints and regular sleep/wake patterns.

Participants entered the sleep laboratory at 21.00h

and were prepared for PSG measurements. Lights

were turned off at around 23.00h. The waking up sig-

nal was given around 7 o’clock.

Sleep recordings and analysis Polysomnographic

sleep recordings were obtained during all sleep

episodes with a digital recorder, and included EEG

(F3/A2, F4/A1, C3/A2, C4/A1, O1/A2, O2/A1)

sampled at 100Hz, electrooculogram (EOG) sam-

pled at 100Hz, electrocardiogram (ECG) sampled

at 500Hz and chin electromyogram (EMG) sam-

pled at 200Hz. Respiratory effort was measured

with chest and abdominal respiratory effort belts at

10Hz. Obtained PSG recording were scored into

sleep stages using 30s epochs according to standard

criteria (Rechtschaffen and A.Kales, 1968) by the

Alice Philips Resperonics system and further proof

checked by human experts.

2.2 SecondDataset

The Second data set contains overnight PSG record-

ings of six young healthy volunteers (mean age 27, 4

males and 2 females). They signed a consent form. In

a screening phase, selection of participants was based

on absence of subjective sleep complaints and regu-

larity of sleep/wake patterns. Screening was based

on two questionnaires: the Sleep Disorders Question-

naire (SDQ) (Douglass et al., 1994) and the Pittsburgh

Sleep Quality Index (PSQI) (Buysse et al., 1989). All

selected participants scored within the normal range

of the PSQI. Moreover, the participant should not

score higher than the cutoff scores on the subscales

narcolepsy, apnea, restless legs, and psychiatry of the

SDQ (Douglass et al., 1994).

Participants entered the sleep laboratory at 21.00h

and were prepared for Polysomnographic measure-

ments. Lights were turned off at around 0.00h. The

waking up signal was given around 7 o’clock.

Polysomnographic sleep recordings were ob-

tained during all sleep episodes with a digital

recorder (Vitaport-3, TEMEC Instruments, Kerkrade,

Netherlands), and included EEG (F3/A2, F4/A1,

C3/A2, C4/A1, O1/A2, O2/A1) obtained with the

Sleep BraiNet system (Jordan NeuroScience, San

Bernardino, CA), electrooculogram (EOG), electro-

cardiogram (ECG) and chin electromyogram (EMG).

Respiratory effort was measured with Pro-Tech chest

and abdominal respiratory effort belts (Pittsburgh,

USA). The signals were recorded at a sampling fre-

quency of 256Hz. Obtained PSG recording were

scored by the Siesta Group’s software system - ”Som-

nolyzer” and further proof checked by Siesta’s ex-

perts.

3 METHODS

3.1 Feature Extraction

In the following two subsections we describe the

data preprocessing and feature extraction methodol-

ogy used for both the RHA and the Facial PSG ap-

proaches.

3.1.1 RHA Features

The raw respiration signal was ﬁrst lowpass ﬁl-

tered (cut-off 0.5Hz) and then analyzed for individ-

ual breaths. Based on a localized min/max ﬁlter, lo-

HEALTHINF 2012 - International Conference on Health Informatics

488

Table 2: List of features in the RHA approach, extracted from respiration, heart and wrist actigraphy signals.

Feature number Heart rate features: Description

1 Mean Average heart rate

2 Median Median heart rate

3 Standard deviation Standard deviation of the heart rate

4 Gradient Inclanation coefﬁcient, shows whether heart rate goes up or down within the epoch

5 Variability Heart rate variability

Actigraphy features:

6 Amount of motion over time Scalar with values between 0 and 1, shows relative amount of motion detected

Respiration frequency features:

7 Mean Average respiration rate

8 Median Median respiration rate

9 Standard deviation Standard deviation of the respiration rate

10 Gradient Inclanation coeﬁciant, shows whether respiration rate goes up or down within the epoch

Respiration amplitude features:

11 Mean Average of amplitudes of respiration cycles

12 Median Median of amplitudes of respiration cycles

13 Standard deviation Standard deviation of amplitudes of respiration cycles

14 Gradient Inclination coefﬁcient, shows whether amplitude goes up or down within the epoch

cal minima and maxima were detected. When found

in the right order, they characterize a single breath.

Based on the distribution of identiﬁed breath ampli-

tudes in a signal, breaths which were too small or too

large (outliers) were removed.

In a similar manner, the ECG signal is lowpass ﬁl-

tered (cut-off 5 Hz) and de-trended. Individual heart

beats are detected using pattern matching. Again, out-

lier removal is applied and the resulting signal is a se-

quence of inter beat intervals (IBIs), which has been

transformed into (instantaneous) heart rate (in bpm)

by taking its reciprocal and multiplying by 60. The

wrist actigraphy signal has been low-passed using a

running average ﬁlter (5x30 seconds epochs in size)

and further normalized on a unit interval.

Figure 1: Respiration rate, heart rate and actigraphy signals

as they change over the night, plotted along the two hypno-

grams from two experts. Colors in the lower graph represent

sleep states.

Sleep is normally scored (Rechtschaffen and

A.Kales, 1968) in non-overlapping 30-second long in-

tervals (epochs). Thus, features of respiration, heart

and actigraphy signals are calculated on a per-epoch

basis. Table 2 lists the complete set of features ex-

tracted from the signals.

3.1.2 Features in the Facial PSG Approach

The raw signal used for feature extraction in the Fa-

cial PSG approach was recorded by electrodes placed

at the following three standardized locations:(1) upper

left eye (EOG Left), (2) left mastoid bone (reference

A1) and (3) ground electrode at the neck of the partic-

ipant. Given this setup for signal extraction the signal

recorded at A1 channel was subtracted from the sig-

nal of the EOG L channel. Furthermore, to estimate

the power spectral density of each epoch, we applied

Welch’s method (Welch, 1967). Figure 2 shows re-

sults of the Welch’s method where the color repre-

sents the power at a certain frequency (top plot).

To facilitate the visual interpretation of the rela-

tion between the Welch’s power plot(features) and

the reference scoring (labels), the bottom plot in the

ﬁgure shows corresponding hypnogram and the mid-

dle plot shows power plot but speciﬁcally for low

frequencies which correspond to deeper sleep (a.k.a.

slow wave sleep, SWS). As it can be easily seen,

peaks of power in SWS plot correspond to N3 sleep

stages of the hypnogram.

For the machine learning part of the Facial

PSG approach input-output pairs were constructed in

the following manner: for each epoch, a power spec-

trum vector was computed and coupled with a sleep

stage label. This resulted in about 800 input-output

pairs per subject (corresponding to 7 hours of sleep).

TOWARDS UNOBTRUSIVE AUTOMATED SLEEP STAGE CLASSIFICATION - Polysomnography using Electrodes

on the Face

489

50Hz power line

artifact

High density of low

frequency, deeper

sleep

stripes are

motions, either

eyes or body

only low frequency

power plot

PSG based,

Hypnogram

-1 -REM

-2 wake

-3 N1 stage

-4 N2 stage

-5 N3 stage

lfa (11Hz), eyes

closed, but not

asleep yet

Figure 2: Facial PSG approach, features and labels (from

top-to bottom): (1) Signal power vs. frequency over

time, (2) Low frequency(deeper sleep) power over time,

(3) hypnogram plot.

3.2 Classiﬁcation

Robust Soft Learning Vector Quantization (RSLVQ)

is one of many LVQ variants, originally developed

by Kohonen (Kohonen, 1995). This family of ma-

chine learning algorithms has been applied to classi-

ﬁcation problems in many ﬁelds (Centre, 2002) and

is characterized by its computational efﬁcacy. LVQ

is a method of prototype-based, multi-class classi-

ﬁcation, where each class is represented by one or

more prototypes. A prototype is deﬁned as a point

in the N-dimensional feature space with an accom-

panying class label, and trained by sequential han-

dling of training data. Each time a training sample is

presented, the closest prototypes with correct and in-

correct label are pulled towards or pushed away from

the training sample, respectively. When training pro-

gresses, the prototypes will progressively better rep-

resent the classes. When applied to unseen data, clas-

siﬁcation is performed by returning the label of the

closest prototype. Usually Euclidean distance is used

as distance measure.

In a recent study (Witoelar et al., 2010), the per-

formance of several LVQ variants in a controlled envi-

ronment was analyzed. The (relative) robustness and

convergent properties (i.e., insensitivity to overtrain-

ing) motivated our choice for RSLVQ, as proposed

in (Seo and Obermayer, 2003). In this ’soft’ version of

LVQ the magnitude of displacement of prototypes in

each training step is relative to their distance from the

training sample. This method makes an assumption

on the distribution of data samples around the pro-

totypes, which we chose to be Gaussian with equal

variances (for each prototype). The total distribution

of data from a single class therefore is assumed to be

a mixture of Gaussian distributions.

3.3 Performance Measurement

In order to allow in depth comparisons of the two

techniques we present classiﬁcation results of both

experiments in the shape of confusion matrixes.

Essentially confusion matrix contains three widely

known (in classiﬁcation tasks assessments) compari-

son entities: (1) confusion matrix, (2) percentage of

agreement and (3) Cohen’s Kappa agreement coef-

ﬁcient. The confusion matrix can be used for de-

tailed assessment of classiﬁer’s performance in terms

of which classes are often mistaken for what other

classes. Furthermore they allow calculation of a base-

line performance based on just class priors. Since we

were mostly interested in overall performance assess-

ment, in section 4 for each cycle of the cross valida-

tion scheme we only present it’s outcome with two

values: (1) percentage of agreement and (2) Cohen’s

Kappa coefﬁcient.

4 RESULTS AND DISCUSSION

This section presents the results obtained by two

sleep monitoring approaches, namely Facial PSG and

RHA. Section 4.1 reports the results obtained with Fa-

cial PSG while Section 4.2 reports the results obtained

with the RHA approach. Both subsections contain

tables presenting percentages of agreement and Co-

hen’s Kappa coefﬁcients per cross validation run, as

well as overall agreement matrices.

4.1 Facial PSG Results

Table 3 shows Cophen’s Kappa and percentage of

agreement ﬁgures per run of the cross-validation

scheme. The last column contains average values.

Table 3: Facial PSG approach results, per cross validation

run (per subject).

Subjects (Second study)

1 2 3 4 5 6 mean

Agreement % 76.20 71.43 71.46 64.71 80.91 82.07 74.46

Cohen’s Kappa 0.66 0.58 0.62 0.50 0.69 0.74 0.63

Table 4 shows the overall agreement matrix that

contains: confusion matrix (in bold), percentage of

agreement, Cohen’s Kappa coefﬁcient, positive pre-

dictive values (PPV) and sensitivity of the classiﬁer

per class.

From table 4 it can be seen that the overall perfor-

mance signiﬁcantly exceeds random guessing, which

is 1989/6292 = 31.61%. Furthermore it can be seen

the largest number of confusions is for actual wake

epochs being (falsely) recognized as light sleep. Ac-

tually, the classiﬁer is falsely biased towards light

HEALTHINF 2012 - International Conference on Health Informatics

490

Table 4: Facial PSG approach overall results. Confusion

matrix (in bold), percentage of agreement, Cohen’s Kappa

coefﬁcient, positive predictive values (PPV) and sensitivity

of the classiﬁer per class.

Overall Wake Light Deep Rem Sum Sensitivity

Wake 1206 42 8 42 1298 92.91%

Light 320 1931 456 430 3137 61.56%

Deep 22 102 735 15 874 84.10%

Rem 26 118 16 796 956 83.26%

Sum 1574 2193 1215 1283 6265

PPV 76.62% 88.05% 60.49% 62.04%

Agreement 74.51%

Cohen’s

Kappa

0.64317

sleep, as it classiﬁes half of the total number of epochs

as light sleep (i.e, 3173/6292 = 50.43%), resulting in

a low sensitivity (52.66%) for that class.

In addition to numerical representation of the

classiﬁcation ﬁgure 3 shows both input data (Facial

PSG spectrum; middle plot) and graphical representa-

tion of the hypnograms both target and estimated (bot-

tom plot). The top plot of the ﬁgure shows the power

spectrum of C4-A1 PSG channel, which served as an

input for an additional experiment we conducted. The

essence of the experiment was in substituting Facial

PSG signal with C4-A1 signal. Given the fact that

C4 electrode of the PSG setup is mounted close to

the brain and subsequently has a stronger signal, our

assumption was to observe gain in classiﬁcation per-

formance. However, this experiment proved an op-

posite effect. Despite better signal to noise ratio, a

signiﬁcant drop in performance of the classiﬁer ap-

peared. This may indicate that when it comes to a

single channel PSG, electrode-positions in the Facial

PSG are better suited for sleep stage estimation.

Figure 3: Results and input data (spectrum) visualization

for subject ID7.

Table 5 shows overall performance matrix for C4-

A1 channel. From this table we may see (when

compared to Table 4) that Cohen’s Kappa statistics

lowered by 0.0662 and percentage of agreement by

6.55%.

Table 5: Facial PSG approach overall results on C4-A1

channel. Confusion matrix (in bold), percentage of agree-

ment, Cohen’s Kappa coefﬁcient, positive predictive values

(PPV) and sensitivity of the classiﬁer per class.

Overall Wake Light Deep Rem Sum Sensitivity

Wake 1199 18 15 66 1298 92.37%

Light 649 1263 799 426 3137 40.26%

Deep 44 97 729 4 874 83.41%

Rem 36 10 241 669 956 69.98%

Sum 1928 1388 1784 1165 6265

PPV 62.19% 90.99% 40.86% 57.42%

Agreement 61.61%

Cohen’s

Kappa

0.49303

4.2 RHA , - Respiration, Heart and

Actigraphy Signals for Hypnogram

Estimation

Table 6 shows Cohen’s Kappa and percentage of

agreement ﬁgures per run of the cross-validation

scheme. The last column contains average values.

Table 6: RHA , - respiration, heart and actigraphy features

approach results, per cross validation run (per subject).

Subjects (Boston data set)

1 2 3 4 5 6 mean

Agreement

48.37 33.62 27.87 32.33 43.74 13.55 33.25

Cohen’s

Kappa

0.26 0.09 0.06 0.09 0.25 0.32 0.18

Table 7 shows the overall agreement matrix that

contains: confusion matrix (in bold), percentage of

agreement, Cohen’s Kappa coefﬁcient, positive pre-

dictive values (PPV) and sensitivity of the classiﬁer

per class.

Table 7: RHA , - respiration, heart and actigraphy features

approach overall results. Confusion matrix(in bold), per-

centage of agreement, Cohen’s Kappa coefﬁcient, positive

predictive values (PPV) and sensitivity of the classiﬁer per

class.

Overall Wake Light Deep Rem Sum Sensitivity

Wake 247 10 8 73 338 73.08%

Light 596 761 723 1020 3100 24.55%

Deep 118 158 370 254 900 41.11%

Rem 221 241 53 368 883 41.68%

Sum 1182 1170 1154 1715 5221

PPV 20.90% 65.04% 32.06% 21.46%

Agreement 33.44%

Cohen’s

Kappa

0.12265

Table 8 shows the agreement ﬁgures presented

earlier in table 1 along with agreement ﬁgures

achieved by RHA and Facial PSG approaches. From

these ﬁgures we notice that the Facial PSG approach

is superior compared to the RHA one in both per-

centages of agreement and Cohen’s Kappa coefﬁcient

numbers. Figures of the RHA approach shows a very

low performance of the classiﬁes when based on res-

piration, heart and actigraphy features. It can be seen

that the overall performance is very close to random

guessing, which is 1715/5221 = 32.85%. Again, the

TOWARDS UNOBTRUSIVE AUTOMATED SLEEP STAGE CLASSIFICATION - Polysomnography using Electrodes

on the Face

491

Table 8: Sleep stages classiﬁcation agreement ﬁgures.

Cohen’s Kappa Agreement(%)

Human expert(PSG) vs. Human ex-

pert (PSG)

0.80 86.00

Automated technique A(PSG) vs.

Automated technique B(PSG)

0.69 85.48

Automated technique(PSG) vs. Fa-

cial PSG

0.64 74.51

Automated technique(PSG) vs.

RHA

0.12 33.44

classiﬁer is falsely biased towards light sleep, as it

classiﬁes most of the total number of epochs as light

sleep (i.e, 3100/5221 = 59.38%), resulting in a very

low sensitivity (24.55%) for that class.

5 CONCLUSIONS

In our study we were not able to ﬁnd signiﬁcant corre-

spondence at individual level(cross subjects) between

PSG based sleep stages estimated by experts and the

features we extracted in the RHA approach. This con-

clusion only holds for this paper’s particular combi-

nation of features and classiﬁer. A separate study on

separability involving at least larger number of sub-

jects is required in order to strengthen this conclusion.

In contrast to the RHA, classiﬁcation, results ob-

tained on features extracted in the Facial PSG ap-

proach look very promising, the good performance

of the classiﬁer indicate good separability of differ-

ent sleep states manifested in Facial PSG features. In

the current study, we employed only ”simple” (low-

capacity, epoch based) classiﬁer that already showed a

performance of 74.51% agreement and 0.64 Cohen’s

Kappa. Performance wise these results position the

Facial PSG approach next to full PSG based auto-

mated techniques (see Table 1). Further improvement

of classiﬁcation performance can be expected when

order and transition probabilities between sleep stages

are taken into account.

In addition to very good performance indicators

the Facial PSG is also much less obtrusive compared

to a full PSG setup. The full PSG setup employs 8+

channels resulting in 12+ electrodes mounted on the

skull of a subject, whereas Facial PSG uses only one

channel (3 electrodes) mounted on a single side of the

face allowing subject to have more natural sleeping

position(s) and what is even more important more nat-

ural sleep. Given its obvious advantage of being more

comfortable the Facial PSG approach makes sleep

studies less labor intensive and consequently more ac-

cessible, also making multiple night studies more re-

alistic which will enable research to better understand

the complex mechanisms involved in humans sleep.

REFERENCES

Buysse, D., Reynolds-III, C., Monk, T., Berman, S., and

Kupfer, D. (1989). Pittsburgh sleep quality index: A

new instrument for psychiatric practice and research.

Psychiatric Research, 28(2):193–213.

Centre, N. N. R. (2002). Bibliography on the self-

organizing maps (som) and learning vector quantiza-

tion (lvq).

Douglass, A., Bornstein, R., Ninomurcia, G., Keenan, S.,

Miles, L., and Zarcone, V. (1994). The Sleep Dis-

orders Questionnaire-I - Creation and Multivariate

Structure of SDQ. Sleep, 17(2):160–167.

Kohonen, T. (1995). Learning vector quantization, The

handbook of brain theory and neural networks. MIT

Press, Cambridge, MA.

Rechtschaffen, A. and A.Kales (1968). A manual of stan-

dardized terminology, techniques and scoring system

for sleep stages of human subjects. U. S. National In-

stitute of Neurological Diseases and Blindness, Neu-

rological Information Network, Bethesda, Md.,.

Seo, S. and Obermayer, K. (2003). Soft learning vector

quantizatio. Neural Computation, 15:1589–1604.

Virkkala, J. (2005). Automatic Sleep Stage Classiﬁcation

Using Electro-oculography. PhD thesis, Faculty of

Computing and Electrical Engineering. Tampere Uni-

versity of Technology.

Welch, P. (1967). The use of fast fourier transform for the

estimation of power spectra: A method based on time

averaging over short, modiﬁed periodograms. IEEE

Transactions on Audio Electroacoustics, AU-15:7073.

Witoelar, A. W., Ghosh, A., de Vries, J. J. G., Hammer, B.,

and Biehl, M. (2010). Window-based example selec-

tion in learning vector quantization. Neural Computa-

tion, 22(11):2924–2961.

HEALTHINF 2012 - International Conference on Health Informatics

492