USING WAVELET TRANSFORM FOR FEATURE EXTRACTION

FROM EEG SIGNAL

Lenka Lhotska, Vaclav Gerla, Jiri Bukartyk

Gerstner Laboratory, Czech Technical University in Prague, Technicka 2, 166 27 Prague 6, Czech Republic

Vladimir Krajca, Svojmil Petranek

University Hospital Bulovka, Budínova 2, 18081 Prague 8, Czech Republic

Keywords: EEG processing, wavelet transform, feature extraction.

Abstract: Manual evaluation of long-term EEG recordings is very tedious, time consuming, and subjective process.

The aims of automated processing are on one side to ease the work of medical doctors and on the other side

to make the evaluation more objective. This paper addresses the problem of computer-assisted sleep staging.

It describes ongoing research in this area. The proposed solution comprises several consecutive steps,

namely EEG signal pre-processing, feature extraction, feature normalization, and application of decision

trees for classification. The work is focused on the feature extraction step that is regarded as the most

important one in the classification process.

1 INTRODUCTION

The electroencephalogram (EEG), describing the

electric activity of the brain, contains a lot of

information about the state of patient health. It has

the advantage of being non-invasive and applicable

over longer time span (up to 24 hours if necessary).

This is an important feature in case we want to

follow disorders that are not permanently present but

appear incidentally (e.g. epileptic seizure) or under

certain conditions (various sleep disorders).

Although the attempts to support EEG evaluation by

automatic or semi-automatic processing have been

made for a long time, there are still many problems

to be solved. We try to contribute by our research to

this effort. The main objective of the described work

is the identification of the most informative features

from sleep EEG records that could be used for

automated (or semi-automated) sleep stage

classification. Our approach to the analysis of

human sleep uses wavelet transform (WT) and

statistics for feature extraction and construction. The

extracted and computed features are used as inputs

for a decision tree (Quinlan, 1990) that is learned to

classify individual sleep stages. We use for our

experiment EEG sleep records rated by an expert,

freely available and downloadable from the Internet

(Kemp, 2007).

The paper is organized as follows. Section 2

describes sleep EEG signal and approaches to its

evaluation. Methods used in our research are

presented in Section 3. Section 4 is devoted to

description of performed experiments. In Section 5

the results of experiments are discussed and the

conclusion is presented in Section 6.

2 SLEEP AND ITS COMPUTER

SUPPORTED CLASSIFICATION

Sleep is a non-uniform biological state that has been

divided into several stages based on

polysomnographic (PSG) measurements that include

EEG, EMG, EOG, ECG, temperature, SpO

(oxygen

saturation of the blood, recorded on the finger),

respiration signals, as well as movement or body

position. Polysomnography is usually performed

over the duration of an entire night, or at least 6.5

hours, in order to investigate normal and disturbed

sleep or vigilance (Bloch, 1997). Normal healthy

sleep is organized into sequences of stages that

typically cycle every 60 – 90 min. The most widely

used standard for terminology and scoring of sleep

236

Lhotska L., Gerla V., Bukartyk J., Krajca V. and Petranek S. (2008).

USING WAVELET TRANSFORM FOR FEATURE EXTRACTION FROM EEG SIGNAL.

In Proceedings of the First International Conference on Bio-inspired Systems and Signal Processing, pages 236-241

DOI: 10.5220/0001062002360241

 SciTePress

stages is the manual by Rechtschaffen and Kales

(RK) (Rechtschaffen and Kales, 1968). A standard

summary method is the hypnogram that graphically

represents sleep stages in 20-30 second epochs. The

PSG can be generally divided into epochs of 10, 20,

30, or 60 s, which are then visually classified into

one of RK stages by a sleep technologist. The

resulting time evolutionary description of sleep in

terms of stages, termed hypnogram, is used by

physicians for diagnosis. The Rechtschaffen and

Kales manual details a complete process of

recording and analysing sleep, which is followed by

the vast majority of sleep laboratories, worldwide.

On the basis of EEG (plus EOG and EMG), epochs

can be scored into sleep stages:

• Stage 1 – shallow/drowsy sleep;

• Stage 2 – light sleep;

• Stage 3 – deepening sleep;

• Stage 4 – deepest sleep;

• Stage REM – dreaming sleep.

Stages 1 to 4 are frequently described as non-

REM sleep, and stages 3 and 4 are described as slow

wave sleep (SWS). Other scores are Wake (W) and

Movement Time (MT). Since the depth of sleep

changes continuously, the artificial demarcation of

sleep stages by the RK classification is a

simplification. The exact time of change of state is

highly subjective and leaves room for interpretation

by the physician who scores transitional epochs

(e.g., Stage 1 and Stage 3) differently on different

occasions (Schaltenbrand, 1996).

Studies have shown agreement between

physicians performing scoring that ranges from 67%

to 91% (Gaillard and Tissot, 1973), (Stanus et al.,

1987), (Kim et al., 1992), depending on different

scoring epoch lengths and number of readers.

However it is necessary to remark that most data on

interscorer agreement are based on the study of

normal subjects. Processing of sleep recordings

requires elaborate training and is time consuming

and expensive. No generally accepted standard

exists for automatic sleep staging, but

computerization can improve efficiency and reduce

cost (Doman, 1995), and enhance collaboration

between laboratories (Kemp, 1993).

Various approaches to computer classification of

PSGs have been used. Johnson et al. (Johnson et al.,

1969) presented a spectral analysis study of the EEG

in different stages, which was subsequently used by

Larsen and Walter (Larsen and Walter, 1970) to

develop an automated staging technique based on

multiple-discriminant analysis. Agarwal and Gotman

(Agarwal and Gotman, 2001) use a method based on

the segmentation and self-organization technique.

The following five steps are necessary to perform

computer-assisted staging: segmentation; feature

extraction; clustering; assignment of stages to

different clusters of patterns; and optional smoothing

of the hypnogram. The study (Agarwal and Gotman,

2001) shows that the greatest discrepancy occurs in

Stage 1. The sensitivity and the specificity are

38.6% and 43.4%, respectively. This is to be

expected in the highly transitional Stage 1. Stage 1

also has significant similarities to REM stage and

can be considered as one stage away from Stage 1.

Moreover, it is accepted that manual scoring of

Stage 1 is the most subjective due to its transitional

nature.

3 METHODS

In our study we have used similar procedure as

Agarwal and Gotman and the same we used in one

of our previous studies (Gerla, Lhotska, and Krajca,

2005). The sleep EEG signal classification

comprises several steps: segmentation, feature

extraction, feature normalization, feature selection,

and generation of decision trees.

We have applied wavelet transform (Daubechies,

1992) to sleep EEG signal preprocessing. Mean of

the signal is calculated and subtracted from a signal

before WT is applied. Discrete Wavelet Transform

(DWT) represented by a filter bank is employed for

wavelet decomposition. Before the decomposition

starts it is necessary to select a mother wavelet used

for defining FIR filters and a level of a

decomposition tree. For deciding which mother

wavelet should be selected we consider the impulse

response and amplitude frequency characteristics of

the FIR filter specified by the corresponding mother

wavelet. After the DWT is done we get

approximation and detail coefficients as input data

for further processing. Then the segmentation is

performed.

Segmentation. The non-adaptive segmentation is

employed. Non-adaptive or constant segmentation

divides a signal into segments of a constant length.

This kind of segmentation is basically the easiest

one. The disadvantage of this method is that the

segments are not necessarily stationary. The length

of a segment is chosen regarding the character of

data.

Feature extraction is the second most important

part after wavelet decomposition. It is a process

which changes representation of segments by

extracting features from them. The aim is to select

those features which carry most information about

the segment. The statistic parameters are in principle

very suitable for this purpose. We use autoregressive

features and computed wavelet coefficients as well.

We use the following parameters: average absolute

USING WAVELET TRANSFORM FOR FEATURE EXTRACTION FROM EEG SIGNAL

237

amplitude, maximal positive amplitude, maximal

negative amplitude, maximal absolute amplitude,

frequency weighted energy, sample mean, sample

central moment, sample variance, statistical median,

energy, and entropy. The autoregressive features are

calculated from the transfer function of an

autoregressive model, in which a present value x

future values x

n+I

, i=1,2,… are estimated by using

the previous values {x

n-m

,…. x

n-1

} (Therrien, 1992).

We can extract features from each source (an

original signal, its first and second derivation)

independently.

Feature normalization. Mean and standard

deviation of extracted features are different. That

could have a negative influence to the classification

process, when a classifier uses distances between

points in n-dimensional space. Before we start

classification the features must be normalized to

have the same mean and standard deviation. The

features have normal distribution N(0,1).

Feature reduction. There are several different ways

in which the dimension of a problem can be reduced.

In this work Principal Component Analysis (PCA)

(Smith, 2002) approach is used which defines new

features (principal components or PCs) as mutually-

orthogonal linear combinations of the original

features.

Feature selection is considered successful if the

dimensionality of the data is reduced and the

accuracy of a learning algorithm improves or

remains the same. Decision tree algorithms such as

C4.5 can sometimes overfit training data, resulting

in large trees. In many cases, removing irrelevant

and redundant information can result in C4.5

producing smaller trees. The Chi-squared statistic is

used for feature selection.

Classification. We have decided to use decision tree

algorithms because they are robust, fast, and what is

important especially in medical domain their results

are very easy to interpret. In particular, the C4.5

algorithm has been applied, namely its J48 variant

available in the Weka software tool (Frank et al.,

2007).

Success rate of classification. As a measure of

success rate we have used the overall accuracy of the

classification. The overall accuracy is calculated as

the relative number of correct decisions.

4 EXPERIMENTS

The main purpose of our experiments has been to

find the most suitable wavelet decomposition and

the most discriminative features to achieve good

classification results. The analyzed EEG recordings

are presented in the next section and then our

experiments with EEG data are described.

4.1 Source of EEG Recordings

Our source of EEG recordings is The Sleep-EDF

Database (Kemp, 2007). Four EEG recordings from

different subjects were downloaded. The recordings

were obtained from Caucasian males and females

(21 - 35 years old) without any medication. They

contain horizontal EOG, Fpz-Cz and Pz-Oz EEG,

each sampled at 100 Hz. The recordings also contain

the submental-EMG envelope, oro-nasal airflow,

rectal body temperature and an event marker, all

sampled at 1 Hz. Hypnograms are also added which

are manually scored according to Rechtschaffen &

Kales based on Fpz-Cz / Pz-Oz EEG instead of C4-

A1 / C3-A2 EEG (Sweden et al., 1990).

Subjects, recordings and hypnogram scoring for

the 4 sc* recordings are described in (Mourtazaev,

1995). Subjects and 4 st* recordings are more

extensively described in (Kemp et al., 2000). The

sleep stages Wake, Stage1, Stage2, Stage3, Stage4,

REM and 'unscored' are coded in the file as binaries

0, 1, 2, 3, 4, 5, 6 and 9.

After reviewing the data we have found out that

the classes in data are unevenly represented. Class 1

(Wake) is the most frequent one and class 5 (stage4)

occurs sporadically. We have generated the training

set in which all classes are equally represented. That

means that a classification error caused by an

unequal distribution of classes should be reduced.

4.2 Experiment 1

A goal of this experiment is to find features which

contain the information about classes included in

data. In other words the features should be highly

correlated with the class. In our case we have six

classes (wake, stage1, stage2, stage3, stage4, REM).

This is a complex task and it is quite impossible to

find only one feature to correlate with all classes.

We modify our goal to examine all features for

every combination of two different classes and select

the most significant feature for discriminating the

classes from each other. There are 15 combinations

so we get 15 features in total. We have chosen EEG

sample (sc4002e0), which includes all 6 classes;

each having 200000 samples (2000 seconds). For

WT, the following setting has been used: level of

decomposition tree 7; mother wavelet db6; wavelet

coefficients used for feature extraction (2,1), (3,1),

(4,1), (5,1), (6,1), (7,1), (7,0); segment length 10s.

The results of this experiment and the best

features selected for classification of every

combination of two different classes are shown in

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

238

Table 1: Results of experiment 1 and the best features selected for differentiation between couples of classes.

Stage Wake Stage 1 Stage 2 Stage 3 Stage 4 REM

class 1 2 3 4 5 6

1 96% - f1 97.5% - f2 99.5% - f3 99.5% - f5 98.9% - f1

2 96% - f1 85% - f7 91.5% - f8 98.5% - f9 70% - f10

3 97.5% - f2 85% - f7 73% - f11 94% - f12 85% - f4

4 99.5% - f3 91.5% - f8 73% - f11 85% -f3 94.5% - f13

5 99.5% - f5 98.5% - f9 94% - f12 85% -f3 99.4% - f6

6 98.9% - f1 70% - f10 85% - f4 94.5% - f13 99.4% - f6

Table 2: Description of the used features.

feature original name of a feature

source for

extraction

wavelet

coefficient

full name of the feature

f1 MeaAbV_1d_d2_Fpz-Cz first derivation D2 (2.1) average absolute amplitude

f2 Energy_sg_d4_Pz-Oz signal D4 (4.1) energy

f3 MeaAbV_Sg_d5_Pz-Oz signal D5 (5.1) average absolute amplitude

f4 Energy_1d_d5_Fpz-Oz first derivation D5 (5.1) energy

f5 Energy_1d_d5_Pz-Oz first derivation D5 (5.1) energy

f6 FrWeiE_Sg_d6_Fpz-Cz signal D6 (6.1) frequency weighted energy

f7 FrWeiE_1d_d5_Pz-Oz first derivation D5 (5.1) frequency weighted energy

f8 FrWeiE_Sg_d5_Pz-Oz signal D5 (5.1) frequency weighted energy

f9 MeaAbV_Sg_d5_Fpz-Cz signal D5 (5.1) average absolute amplitude

f10 MeaAbV_Sg_d3_Pz-Oz signal D3 (3.1) average absolute amplitude

f11 FrWeiE_Sg_d7_Pz-Oz signal D7 (7.1) frequency weighted energy

f12 Energy_Sg_d6_Pz-Oz signal D6 (6.1) energy

f13 Energy_1d_d6_Fpz-Cz first derivation D6 (6.1) energy

Table 1. The names and sources of these features are

presented in Table 2. There are five classification

results below 90% as it is shown in Table 1. It

means that we are not able to extract any single

feature which can separate these particular

combinations of two classes. There must be used a

combination of features. We can see that there are

two features (f1, f3) occurring not only once as most

discriminative. Each of them is chosen to be the

discriminative feature for two combinations. A set of

features is therefore reduced and we have 13

features. Unfortunately 5 of the features (marked in

italics in Table 1) are not good enough for

classification and thus we have decided to perform

another experiment where different wavelet

coefficients are decomposed and other features are

examined.

4.3 Experiment 2

The goal is implicated by the result of the previous

experiment. There are five combinations of two

classes (4x3, 6x2, 3x2, 5x4, 6x3) which are

classified with success rate lower than 90% by using

features extracted from the wavelet coefficients.

Now we try to achieve more accurate results by

extracting features from such wavelet coefficient

that have the same frequency resolution. Frequency

resolution of a wavelet coefficient depends on

sample frequency of data (100Hz) and on the level

of the wavelet coefficient. We may be able to find

more specific features carrying more information

about separability of classes. Two different settings

and wavelet decomposition trees are used: 1. level of

decomposition tree 4; mother wavelet db15; wavelet

coefficients used for feature extraction

(4,0), (4,1),

(4,2), (4,3), (4,4), (4,5), (4,6), (4,7)

; segment length 10s;

2. level of decomposition tree 5; mother wavelet

db20; wavelet coefficients used for feature

extraction

(5,0), (5,1), (5,2), (5,3), (5,4), (5,5), (5,6),

(5,7), (5,8), (5,9), (5,10), (5,11), (5,12), (5,13), (5,14),

(5,15)

; segment length 10s. Wavelet coefficients

from the highest level of decomposition trees are

used for feature extraction. They have the highest

frequency resolution compared with others in the

wavelet decomposition tree. We can assume that the

features extracted from these coefficients carry

different piece of information about classes.

The results of this experiment have not been so

successful as we have expected. The only feature

that has brought relatively significant improvement

of differentiation between two classes (by 7.5%) has

been average absolute amplitude FPz-Cz (wavelet

USING WAVELET TRANSFORM FOR FEATURE EXTRACTION FROM EEG SIGNAL

239

coefficient (5,1)). The experiment has shown that

application of other mother wavelets (db20 or

higher) and different wavelet decomposition trees

could result in finding new more discriminative

features.

4.4 Experiment 3

The final experiment has been divided into three

parts, namely using different groups of classes. EEG

recordings (sc4012e0, st7022j0, sc4102e0) have

been used as testing sets for this purpose.

Part 1. We have classified data into all six classes

using all features as described in experiment 1. The

results have verified our assumption that the features

f3, f4, f7, f10 and f11 which do not separate classes

well (see Table 1) decrease final classification

accuracy.

Part 2. Based on the experiment 1 we have tried

to distinguish among four classes, organized in two

groups, namely (1, 3, 5, 6) and (1, 4, 5, 6). We have

used six features from the original 15 for each

group. The classification results for the first group

have been negatively influenced by the feature f3

and for the second group by the feature f4.

Part 3. We have verified well discriminating

features discovered in experiment 1. For this

purpose we have selected three groups of three

classes each that can be separated very well by these

features. The three groups are composed of the

following classes (1, 5, 6), (1, 2, 5) and (1, 4, 6).

All results are summarized in Table 3. The record

sc4102e0 has not been used in those experiments

where class 5 has been tested because it does not

contain any segment belonging to class 5. The

classification results are as we have assumed. They

are mainly affected by low discriminability between

classes 2 (stage1) and 3 (stage2) and classes 2

(stage1) and 6 (REM).

Table 3: Results of experiment 3 (success rate of

classification).

classes sc4002e0 sc4012e0 st7121j0 sc4102e0

1,2,3,4,5,6 72.1% 69.5% 63% x

1,3,5,6 87.2% 87.6% 78.4% x

1,4,5,6 87.8% 82.9% 73.2% x

1,5,6 98.3% 94.5% 92% x

1,2,5 97.2 92% 87% x

1,4,6 96.5% 91.3% 90% 87%

5 DISCUSSION

Tables 1 and 2 are used for the interpretation. When

we look at Table 2 we can see that all features

extracted for the classification task in the experiment

1 are based on energy, mean absolute amplitude and

frequency weighted energy. These features reflect

the changes of energy in the given wavelet

coefficient which is related to a specific frequency

spectrum. This is very important as we see later.

Now we try to explain why we have got the results

of classification shown in Table 1. When we look at

this table we can see that successful classifications

are for the classes classified with features extracted

from wavelet coefficients which have the frequency

spectrum same as the frequency spectrum only a

single class in the set of two classes has. That means

that the feature used for such classification has high

energy for this class and small energy for the other

one. Then we can simply use a threshold to separate

these two classes from each other. When we look at

Table 1 again we can see that all successful

classification results (success rate higher than 90%)

are achieved between classes with mutual distance

more than one class, for example, classes 1x3, 1x4,

2x5 etc. It is because the distance between these

classes is quite long which is required for successful

classification. An exception is the class 1 which is

classified correctly in all cases. When we examine

frequency spectra of classes 1 (Wake) and 2

(stage1), we find out that they are well separable.

However we have to note that there exists

overlapping (some frequencies occur in

neighbouring stages). Therefore poorer classification

result (below 90%) is for classes just next to each

other (2x3, 3x4, 4x5 and 6x2). Unfortunately we

have not yet found any feature better describing the

classes by using different wavelet decomposition

(experiment 2). The results of classification in

experiment 3 are affected by this fact as well. In the

following paragraph we suggest some ideas which

could improve classification of sleep EEG data.

The frequency resolution of wavelet coefficients

in level 5 of a wavelet decomposition tree is 3.12Hz.

This decomposition is used in experiment 2. It was

not detailed enough for distinguishing incorrectly

classified classes. So we propose to make the

frequency resolution higher by getting wavelet

coefficients from level 6 (frequency resolution

1.57Hz) or even level 7 (0.78Hz). For these purposes

we must ensure that the filter used for

decomposition has steep frequency characteristic.

We would recommend to use mother wavelets db20

and higher. If this condition is satisfied the results

would not be influenced by leakage of other

frequency components (antialiasing).

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

240

6 CONCLUSIONS AND FUTURE

WORK

Sleep problems belong to the most common serious

neurological disorders. Reliable and robust detection

of these disorders would improve the quality of life

of many people. The implemented methods allow

automatic classification of EEG signals. The

approach has been tested on real sleep EEG

recording for which the classification has been

known. We have focused on discovering the most

significant features which would be highly

correlated with classes of data. Our experiments

have been based on the selection of a single feature

to separate data belonging to two classes. There have

been many other features with good selection

results. The most frequent ones have been

autoregressive features representing the order of

used AR model and error of AR model. We have

determined some features and wavelet coefficients

which are best suited for classification of sleep EEG

data. The future work will be focused on

exploitation of other types of mother wavelets, using

higher level of wavelet coefficients as source of

features, and more sophisticated classifiers.

ACKNOWLEDGEMENTS

This work has been supported by the research

program “Information Society” under Grant No.

1ET101210512 “Intelligent methods for evaluation

of long-term EEG recordings“.

REFERENCES

Agarwal, R., Gotman, J., 2001. Computer-assisted sleep

staging, IEEE Trans. on Biomed. Engineering, 48

(12), pp. 1412-1423.

Bloch,

K.E., 1997. Polysomnography: a systematic review,

Technology and Health Care, 5, pp. 285-305.

Daubechies, I., 1992. Ten lectures on Wavelets, CBMS-

NSF, SIAM, 61, Philadelphia, Pennsylvania, USA.

Doman, J., Detka, C., Hoffman, T., Kesicki, D., Monahan,

J.P., Buysse, D.J., Reynolds III, C.F., Coble, P.A.,

Matzzie, J., Kupfer, D.J., 1995. Automating the sleep

laboratory: implementation and validation of digital

recording and analysis, Int. J. Biomed. Comput., 38,

pp. 277-290.

Frank, E., Hall, M., Trigg, L. Weka – Data Mining

Software in Java. Internet site address: http://

www.cs.waikato.ac.nz.

Gaillard, J.M., Tissot, R., 1973. Principles of automatic

analysis of sleep records with a hybrid system,

Comput. Biomed. Res., 6, pp. 1-13.

Gerla, V., Lhotská, L., Krajča, V., 2005. Utilization of

Time Dependence in EEG Signal Classification. The

3rd European Medical and Biological Engineering

Conference EMBEC´05 Prague.

Johnson, L., Lubin, A., Naitoh, P., Nute, C., Austin, M.,

1969. Spectral analysis of the EEG of dominant and

nondominant alpha subjects during waking and

sleeping, Electroencephalogr. Clin. Neurophysiol., 26,

pp. 361-370.

Kemp,

B., 1993. A proposal for computer-based

sleep/wake analysis, J. Sleep Res., 2, pp. 179-185.

Kemp, B., Zwinderman, A., Tuk, B., Kamphuisen, H.,

Oberyé, J., 2000. Analysis of a sleep-dependent

neuronal feedback loop: the slow-wave

microcontinuity of the EEG. IEEE-BME 47(9), pp.

1185-1194.

Kemp, B., 2007. Sleep Recordings and Hypnograms in

European Data Format. The Netherlands. Internet site

address: http://www.physionet.org/ physiobank/

database/sleep-edf.

Kim, Y., Kurachi, M., Horita, M., Matsuura, K.,

Kamikawa, Y., 1992. Agreement in visual scoring of

sleep stages among laboratories in Japan, J. Sleep Res.,

pp. 58-60.

Larsen, L.E., Walter,

D.O., 1970. On automatic methods

of sleep staging by EEG spectra, Electroencephalogr.

Clin. Neurophysiol., 28, pp. 459-467.

Mourtazaev, M., Kemp, B., Zwinderman, A.,

Kamphuisen, H., 1995. Age and gender affect

different characteristics of slow waves in the sleep

EEG. Sleep 18(7), pp. 557-564.

Quinlan, J.R., 1990. Decision trees and decision making.

IEEE Trans System, Man and Cybernetics, 20(2), pp.

339-346.

Rechtschaffen, A., Kales, A. (eds.), 1968. A manual of

standardized terminology, techniques and scoring

system for sleep stages of human subjects, Brain

Inform. Service/Brain Res. Inst., Univ. California, Los

Angeles.

Schaltenbrand,

N. et al., 1996. Sleep stage scoring using

the neural network model: Comparison between visual

and automatic analysis in normal subjects and patients,

Sleep, 19, pp. 26-35.

Smith, L.I., 2002. A tutorial on Principal Components

Analysis, University of Otago, New Zealand, 2002.

Stanus, E., Lacroix, B., Kerkhofs, M., Mendlewicz, J.,

1987. Automated sleep scoring: A comparative

reliability study of algorithms, Electroencephalogr.

Clin. Neurophysiol., 66, pp. 448-456.

Sweden, B., Kemp, B., Kamphuisen, H., Velde, E., 1990.

Alternative electrode placement in (automatic) sleep

scoring (Fpz-Cz / Pz-Oz versus C4-A1 / C3-A2). Sleep

13(3), pp.279-283.

Therrien, C.W., 1992. Discrete Random Signals and

Statistical Signal Processing. Englewood Cliffs (NJ):

Prentice Hall.

USING WAVELET TRANSFORM FOR FEATURE EXTRACTION FROM EEG SIGNAL

241