FEATURE EXTRACTION AND SELECTION FOR AUTOMATIC

SLEEP STAGING USING EEG

Hugo Simões, Gabriel Pires

Institute of Systems and Robotics, University of Coimbra, Coimbra, Portugal

Urbano Nunes, Vitor Silva

Department of Electrical Engineering, University of Coimbra – Polo II, Coimbra, Portugal

Keywords: Feature Extraction, Feature Selection, EEG Sleep Staging, Bayesian Classifier.

Abstract: Sleep disorders affect a great percentage of the population. The diagnostic of these disorders is usually made

by a polysomnography, requiring patient’s hospitalization. Low cost ambulatory diagnostic devices can in

certain cases be used, especially when there is no need of a full or rigorous sleep staging. In this paper,

several methods to extract features from 6 EEG channels are described in order to evaluate their

performance. The features are selected using the R-square Pearson correlation coefficient (Guyon and

Elisseeff, 2003), providing this way a Bayesian classifier with the most discriminative features. The results

demonstrate the effectiveness of the methods to discriminate several sleep stages, and ranks the several

feature extraction methods. The best discrimination was achieved for relative spectral power, slow wave

index, harmonic parameters and Hjorth parameters.

1 INTRODUCTION

About a third of the population suffers from sleep

disorders, including the obstructive sleep apnea

syndrome (Doroshenkov et al, 2007). The diagnosis

of such diseases is performed by a polysomnography

(PSG) which requires the patient's hospitalization

with costs and discomfort for the patient.

Ambulatory diagnostic devices may have an

important role in order to mitigate these factors. The

PSG consists on the acquisition of various electrical

biosignals including electroencephalogram (EEG),

electrooculogram (EOG) and electromyogram

(EMG). The signals are segmented into epochs of 30

seconds and assigned to a sleep stage by an expert

(Iber et al, 2007). This is a tedious and time

consuming task. Automatic sleep stages

classification (ASSC) is therefore an attractive

solution. However, the general opinion is that most

of the experts do not rely on ASSC software,

because they usually present a low performance (i.e.

present a high level of disagreement). One of the

main reasons is due to the high variability between

subjects which makes it difficult to obtain robust

models for classification. The expert uses sometimes

heuristics difficult to implement in the algorithms

and combines a macro and micro perspective of the

overall epochs. It should be highlighted that there is

also some level of disagreement between experts.

This work describes part of an apnea detection

system to be used in ambulatory situations by

patients at home. It does not intend to substitute the

PSG, but only to determine primarily if the patient is

sleeping at the occurrence of the apnea episode, and

secondly to determine in which sleeping stage it did

occur. The stage classification relies only on EEG

signals. This paper investigates several feature

extraction methods to compare their performance

aiming to achieve improved results in the following

sleep detection stages: wake (W) vs. sleep (S),

NREM (NR) sleep vs. REM (R) sleep, NREM N1

vs. NREM N2 + NREM N3, NREM N1 + NREM

N2 vs. NREM N3, NREM N1 vs. NREM N2,

NREM N2 vs. NREM N3 and NREM N1 vs. REM

sleep (Iber et al, 2007). Moreover, a feature

selection method based on the squared Pearson

correlation coefficient (Guyon and Elisseeff, 2003),

henceforth designated R-square criteria, is applied

with the purpose of finding a reduced set of

discriminative features. These features are used to

provide additional information to the expert, and

also to automatically classify each sleep stage with

some degree of certainty. The classification is

128

Simões H., Pires G., Nunes U. and Silva V. (2010).

FEATURE EXTRACTION AND SELECTION FOR AUTOMATIC SLEEP STAGING USING EEG.

In Proceedings of the 7th International Conference on Informatics in Control, Automation and Robotics, pages 128-133

DOI: 10.5220/0002950601280133

 SciTePress

Figure 1: Classification methodology.

performed by a Bayesian classifier using 2-class

detection. Scoring sleep is done according to rules of

the American Academy of Sleep Medicine (AASM)

Manual for Scoring Sleep (Iber et al, 2007), an

actualization of the rules of Rechtschaffen and Kales

(Rechtschaffen and Kales, 1968). According to

AASM Manual, sleep is divided into five stages:

wake, NREM (Non Rapid Eye Movement) sleep

(N1, N2 and N3) and REM (Rapid Eye Movement)

sleep. Considering only EEG signals, the wake stage

is characterized by a low amplitude alpha activity

(8-13 Hz); N1 by a low amplitude theta activity (3-7

Hz); in N2 the predominant frequencies are in the

0.7-4 Hz range and there is the arising of sleep

spindles and K-complexes; N3 presents at least 20%

of the epochs with delta activity (<2 Hz) with

amplitude greater than 75 µV; REM is characterized

by frequencies mostly between 2 and 6 Hz with low

amplitude. Sleep staging based only on EEG

presents some difficulties because different stages

such as wake, REM and NREM N1 present similar

patterns. The ASSC has been addressed by many

research groups. In (Tang et al, 2007), Hilbert-Hang

transform and wavelet transform were applied to

extract harmonic parameters from EEG signals,

(Hese et al, 2001) implemented a semi-automatic

method based on k-means clustering algorithm.

(Ebrahimi et al, 2008) used neuronal networks and

wavelet packet coefficients to discriminate between

different sleep stages. Doroshenkov et al. (2007)

have developed a classification algorithm based on

Hidden Markov Models using only EEG signals.

(Zoubek et al, 2007) have used feature selection

algorithms to find the relevant features extracted

from PSG signals. Schwaibold et al (2003) have

implemented a neuro-fuzzy algorithm to model the

rules of Rechtschaffen and Kales. Although some

studies show good performance, they are very

limited to specific groups of patients and it has not

been possible yet to create generalized models that

provide results accepted by the experts. Moreover, it

remains difficult to discriminate between certain

sleep stages using only EEG signals.

2 DATABASE

Data from all-night PSG records were provided by

the Laboratory of Sleep from Centro Hospitalar de

Coimbra. The PSG was recorded by the model

Somnostar Pro from Viasys at a sampling frequency

of 200 Hz. The database comprises seven patients

(five males and two females) with ages between 27

and 64 years old (mean = 50 years; standard

deviation = 12.88 years). Only six EEG channels

were used: F3-A2, C3-A2, O1-A2, F4-A1, C4-A1

and O2-A1. All recordings were segmented into

epochs of 30 seconds and labelled by an expert.

The dataset was initially composed by 6558

epochs. In order to avoid the over-fitting in the

learning and testing of algorithms, the number of

sleep epochs in the database was reduced to 3000,

balancing the distribution of epochs of different

sleep stages according to a normal night sleep

distribution as presented in Table 1. Since the sleep

stages N2 and N1 are the ones with the highest and

lowest occurrence during a normal night sleep,

respectively, they were set as the stages with major

and minor number of epochs in the dataset,

respectively, and the other sleep stages have a

number of epochs between these limits.

Table 1: Full and reduced datasets.

Sleep

Stages

NREM

Wake N1 N2 N3 REM

Full

dataset

1293 784 2431 1154 896

Reduced

dataset

560 410 760 520 750

3 AUTOMATIC SLEEP SCORING

The classification methodology is illustrated in the

block diagram presented in figure 1. The EEG

signals are filtered and segmented. Different types of

features extraction are used. These features are

EEG Signal

Patient Database

Feature Extraction

Pre

-Processing

Classification

Patient Testing

Notch filter

Butterworth filter

Segmentation

RSP

SWI

Harmonic Parameters

Parameters of Hjorth

Entropy

Skweness & Kurtosis

Feature Selection

R-square

Bayesian Classifier

LDA

Decision Tree

FEATURE EXTRACTION AND SELECTION FOR AUTOMATIC SLEEP STAGING USING EEG

129

then selected using the correlation criteria R-square

measure in order to provide the classification stage,

a Bayesian-based classifier, with the most

discriminative ones. The training process uses data

from a pool of patients and some data from the

patient being monitored, namely, the wake recorded

epochs before the patient fall asleep. This way, the

wake model can be improved. Moreover, the wake

epochs can be used for calibration of sleep stages.

The performance analysis of the of feature extraction

algorithms was done through ten-fold cross

validation. The patients’ database is partitioned into

ten groups with the same number of epochs from

each sleep stage. Nine of them are used to perform

the models of classification and one for testing. This

process is repeated 10 times using a different group

for testing.

4 FEATURE EXTRACTION AND

SELECTION

In ASSC, the EEG is traditionally analyzed in

frequency domain because, according with AASM

Manual, each sleep stage is essentially distinguished

by some spectral properties. However, temporal

analysis provides also useful information. For each

EEG channel, 34 features were extracted using

several methods as described in the following.

Spectral analysis provides some of the most

important features. For each sleep epoch, an

autoregressive method solved by the Yule-Walker

algorithm was applied to estimate the power spectral

density (PSD) (Yilmaz et al, 2007). The spectrum is

divided into ten frequency sub-bands as represented

in Table 2.

Table 2: Spectral sub-bands used in RSP computation.

Bands Sub-bands

Bandwidth

} (Hz)

Delta

Delta 1 {0.5,2.0}

Delta 2 {2.0,4.0}

Theta

Theta 1 {4.0,6.0}

Theta 2 {6.0,8.0}

Alpha

Alpha 1 {8.0,10.0}

Alpha 2 {10.0,12.0}

Sigma

Sigma 1 {12.0,14.0}

Sigma 2 {14.0,16.0}

Beta

Beta 1 {16.0,25.0}

Beta 2 {25.0,35.0}

For each sub-band, the relative spectral power (RSP)

was computed. This parameter is given by the ratio

between the sub-band spectral power (BSP) and the

total spectral power, i.e., the sum of all 10 BSP sub-

bands. This normalization is important to increase

classification robustness during the recording

session.

Some spectral bands can be highlighted over

slow wave bands by means of slow wave index

(SWI) defined by the following ratios:

/( )

elta Theta Alpha

DSI BSP BSP BSP

(1)

/( )

Theta Delta Alpha

TSI BSP BSP BSP

(2)

/( )

Alpha Delta Theta

ASI BSP BSP BSP ,

(3)

where DSI, TSI and ASI stand for delta-slow-wave

index, theta-slow-wave index and alpha-slow-wave

index, respectively (Agarwal et al, 2001).

Harmonic parameters allow the analysis of a

specific band in the EEG spectrum. They include

three parameters: center frequency (f

), bandwidth

) and spectral value at center frequency (S

defined as follows (Tang et al, 2007):

() ()

∑∑

xxc

fPffPf

(4)

( ) () ()

⎟

⎠

⎞

⎜

⎝

⎛

−=

∑∑

xxc

fPfPfff

(5)

(

)

cxxf

fPS

(6)

where, P

(f) denotes the PSD, which is calculated

for the frequency bands {f

} (see Table 2).

The Hjorth parameters provide dynamic

temporal information of the EEG signal.

Considering the epoch x, the Hjorth parameters are

computed from the variance of x, var(x), and the first

and second derivatives x’, x’’ according to (Ansari-

Asl et al, 2007)

)var(xActivity

(7)

)var()'var( xxMobility =

(8)

)'var()var()''var( xxxComplexity ×= .

(9)

The entropy gives a measure of signal disorder

and can provide relevant information in the detection

of some sleep disturbs. It is computed from

histogram of the EEG samples of each sleep epoch,

according with (Zoubek et al, 2007)

ICINCO 2010 - 7th International Conference on Informatics in Control, Automation and Robotics

130

⎛⎞

=−

⎜⎟

⎝⎠

∑

Entropy

(10)

where n is the number of samples within the sleep

epoch, N is the number of bins used in computation

of histogram and n

is the number of samples within

the ith bin.

The skewness is a measure of symmetry. The

kurtosis is a measure of wether the data are peaked

or flat relative to a normal distribution. Defining the

kth order moment m

as (Zoubek et al, 2007)

()

=−

∑

myiy

(11)

where n is the number of samples of an epoch and

is the mean of these samples, the skewness and

kurtosis are given by

223

mmmskewness ×=

(12)

and

224

mmmkurtosis ×= .

(13)

Features are usually selected by wrapper or filter

methods using sequential approaches. The results

from wrappers methods are dependent of the choice

of the classification algorithm. Our option fell on an

R-square filter approach which is independent of the

classifier, based on the Pearson correlation

coefficient defined as (Guyon and Elisseeff, 2003):

()

() ()

varvar

,cov

=ℜ ,

(14)

where X and Y represent two random distributions of

samples, and cov and var designates covariance and

variance, respectively. Considering x

and y

as the

sample values of feature i labelled with class 1 and

class 2, respectively, the value R(i) for the feature i

is given by:

()()

()

()()

ik i ik i

xxyy

xyy

−−

∑

∑∑

(15)

where

and

represent the mean value of x

and

of the m samples. The R-square, computed as

R(i)

, provide a level of discrimination between the

two classes. High values of R-square indicate large

inter-class separation and small within-class

variance. The R-square provides a feature

discrimination ranking.

5 BAYESIAN CLASSIFICATION

The conditional density function of the class i is

modelled as a multivariate distribution under

gaussian assumption

( ) ()()

(

)

|, exp /2

ii i i i

PY K Y Y

μμμ

−

Σ= − − Σ − ,

(16)

where,

()

(

)

K Σ=

(17)

Y is the

feature vector resulting from concatenation

of the extracted features, µ

and Σ

are respectively,

the mean and covariance matrices computed for each

class w

from the training data. The Bayes decision

function is written as:

(

)

(

)( )

{

}

{

()(){}}

221

112

,|maxarg

wPwYp

wPwYpYw

Δ=

(18)

where P(w

) is the ith class prior probability and Δ

an adjustment parameter to control the rate of false

positives and false negatives (Heijden et al, 2004).

6 RESULTS AND DISCUSSION

The feature extraction process provides a vector of

204 features, 34 features per each EEG channel: 10

RSP, 3 SWI, 15 harmonic parameters, 3 Hjorth

Parameters, 1 entropy feature, 1 skewness and 1

kurtosis. Next, the features are sorted in a decreasing

order of level of discrimination by applying the R-

squared based selection approach. Figure 2 shows

the percentage of disagreement for wake/sleep

detection between our ASSC system and expert

classification (i.e. the percentage of epochs for

which the automatic classification differs from

manual classification made by the expert), as

function of the number of features, i. e., the n-most

discriminative features with n = 1,…,52. The

disagreement values are obtained from a ten-fold

cross validation. The lowest disagreement value was

reached using the first 19 ranked features. Table 3

presents the results for each binary classifier, using

1, 2, 3, 19 most discriminative features and all 204

features. Selecting the relevant features reduces the

number of features used in the ASSC leading to an

increased robustness of the classifiers.

The feature selection also enables to identify the

type of features and channels that lead to higher

discrimination results for each 2-class discriminator

FEATURE EXTRACTION AND SELECTION FOR AUTOMATIC SLEEP STAGING USING EEG

131

5 10 15 20 25 30 35 40 45 50

Number of Features

Pe rcenta ge of Disagreement

Wake vs. Sleep

Figure 2: Percentage of disagreement vs. number of

features used in wake vs. sleep classification.

Table 3: Percentage of disagreement obtained using 1, 2, 3

and the 19 most discriminative features and all 204

features.

1 2 3 19 204

W vs. S

11,4 10,7 8,8 7,0 16,7

R vs NR

22,5 21,4 19,5 15,6 30,8

N1 vs. N2/N3

15,1 15,7 15,7 10,6 72,5

N1/N2 vs. N3

15,7 14,7 14,6 15,5 30,3

N1 vs. N2

21,9 22,6 18,5 15,6 63,9

N2 vs. N3

19,0 18,2 16,7 17,7 39,8

N1 vs. R

25,5 24,7 24,4 25,0 64,7

Mean

18,7 18,3 16,9 15,3 45,5

(Table 4). As it can be seen, the feature entropy

(Ent), Skewness (Skw) and kurtosis (Krt) never

appear in the 20 most discriminative features. On the

other hand, the most frequents are the RSP and

harmonic parameters. Analyzing the origin of the 20

most discriminative features for each case, the

parameters of Hjorth (PHj) are most evident in

N1/N2 vs. N3 and N2 vs. N3, but they have no

weight in R vs. NR and N1 vs. R. The harmonic

parameters are more frequent in W vs. S, N1 vs.

N2/N3 and N1 vs. N2, but are not relevant in R vs.

NR, N1 vs. N2/N3, N2 vs. N3 and N1 vs. R. For the

RSP and SWI, they have a similar number of

features in all discriminations, except for N1 vs. R,

where the RSP has several features with good

discrimination, and for N1 vs. N2, where SWI does

not assume any importance. Analyzing the EEG

channels, it can be seen that O1A2 (O1) and O2A1

(O2) are the most relevant in discrimination wake

vs. sleep; F3A2 (F3) and F4A1 (F4) in REM vs.

NREM; and C3A2 (C3) and C4A1 (C4) in N2 vs.

N3. In the remaining discriminations, they all have a

relatively uniform distribution, except in N1 vs. R,

in which the channels O1A2 and O2A1 do not have

any type of contribution. Figure 3 shows the type of

features and channels that lead to higher

discrimination results, taking all discriminators

together. Summarizing, the best ranked

discriminative features never include entropy

features, skewness or kurtosis. These parameters are

related to the signal shape. However, since the EEG

signal patterns are very random, it is difficult to

obtain useful information from these parameters.

Instead, the set of most discriminatory features

between sleep stages was composed mainly by RSP

and Harmonic Parameters. This result emphasizes

the fact that the spectral analysis has more

discriminative information than temporal signal

analysis as already concluded in (Hese et al, 2001;

Tang et al, 2007).

Table 4: Number of feature type and channels within the

20 most discriminative features.

W vs. S

R vs. NR

N1 vs. N2/N3

N1/N2 vs. N3

N1 vs. N2

N2 vs. N3

N1 vs. R

Total

Features

RSP

5 4 6 6 5 6 13 45

SWI

3 2 2 2 0 4 4 17

9 14 8 3 12 2 3 51

PHj

3 0 4 9 3 8 0 27

Ent

0 0 0 0 0 0 0 0

Skw

0 0 0 0 0 0 0 0

Krt

0 0 0 0 0 0 0 0

Channels

1 6 6 3 4 2 5 27

1 3 4 5 4 5 5 27

6 1 3 4 2 4 0 20

2 5 3 2 4 1 5 22

5 3 3 4 4 5 5 29

5 2 1 2 2 3 0 15

RSP SWI HP PHj Ent Sk Krt F3 C3 O1 F4 C4 O2

Features Channels

Number of times it appears

Figure 3: Number of times that each group of features and

each channel appears in the 20 most discriminative

features.

On the other hand, all the 6-six EEG channels

provide useful features for sleep staging

discrimination. Analyzing the results for each of the

binary classifiers, there is greater disagreement in the

case of N1 vs. R sleep. This situation relates to the

fact that, in terms of EEG, the patterns presented in

these two stages are very similar. Finally, a decision

tree was implemented based on 2-class detection, as

represented in Figure 4. At each step, a new level

was introduced from a wake/sleep to all stages

ICINCO 2010 - 7th International Conference on Informatics in Control, Automation and Robotics

132

classification. The results were compared with and

without feature selection (Table 5). The

improvements from feature selection are evident. The

results obtained with our ASSC system are

comparable to the ones obtained in other methods

based on EEG only described in literature (zoubek et

al, 2007; Doroshenkov et al, 2007).

Figure 4: Decision tree based on 2-class detection.

Table 5: Disagreement obtained with using 19 most

discriminative features and all 204 in 2, 3, 4 and 5 sleep

stages classification.

Classification

Diasagreement (%)

All Features 19

2 Class

36 7

3 Class

62 18

4 Class

83 22

5 Class

83 29

7 CONCLUSIONS

In this paper, the use of several feature extraction

methods was investigated in the context of EEG-

based sleep staging. The first conclusion was that the

most discriminative features were determined by

RSP, SWI, Harmonic Parameters and Parameters of

Hjorth. All the 6-EEG channels provide useful

information. On the other hand, the application of

the feature selection method improved, in general,

the process of discrimination by selecting the set of

features that provided a lower percentage of

disagreement. One of the biggest problems in

automatic sleep staging based on EEG is the

similarity between patterns of different sleep stages

such as REM and NREM N1. This can be improved

recurring to other biosignals, such as EOG and

EMG. Another problem in ASSC is the high level of

variability between patients. Using an ambulatory

system, the patient can perform periodic recordings

at home. This way, the first session can be fully

analysed by the expert. The labelled data can be

used to obtain classification models specific to the

patient. Further sessions can then use these robust

user-dependent models. This approach is

under research presently.

REFERENCES

Ansari-Asl. K., Chanel, G., Pun, T., A channel Selection

Method for EEG Classification in Emotion

Assessment Based on Synchronization Likelihood. In

EUSIPCO’07, 1241-1245.

Doroshenkov, L., Konyshev, V., Selishchev, S., 2007,

Classification of Human Sleep Stages Based on EEG

Processing Using Hidden Markov Models. Biomedical

Engineering, 41(1), 25-28.

Ebrahimi, F., Mikaeili, M., Estrada, E., Nazeran, H., 2008,

Automatic Sleep Stage Classification Based on EEG

Signals by Using Neural Networks and Wavelet

Packet Coefficients. In IEEE EMBS’08, 1, 1151-1154.

Guyon, I., Elisseeff, A., 2003, An introduction to variable

and feature selection. Journal of Machine Learning

Research, 3, 1157-1182.

Heijden, F., Duin, R., Ridder, D., Tax, D., 2004,

Classification, Parameter Estimation and State

Estimation. John Wiley & Sons.

Hese, P., Philips, W., Koninck, J., Walle, R., Lemahieu, I.,

2001, Automatic Detection of Sleep Stages Using the

EEG. In IEEE EMBS’01,Proc. of, 1994-1947.

Iber., C., Ancoli-Israel, S., Chesson, A., Quan, S., 2007,

The AASM Manual for the scoring of Sleep and

Associated Events: Rules. Terminology and Technical

Specifications (1st ed.). Westchester, Illinois:

American Academy of Sleep Medicine.

Rechtscheffen, A., Kales, A., 1968, A Manual of

Standardized Terminology, Techniques ans Scoring

System for Sleep Stages of Human Subjects, US

Government Printing Office, National Institute of

Health Publications, Washington DC.

Schwaibold, M., Harms, R., Scholer, B., Pinnow, I.,

Cassel, W., Penzel, T., Becker, H., Bolz, A., 2003,

Knowledge-Based Automatic Sleep-Stage Recognition

– Reduction in the Interpretation Variability.

Somnologie, 7, 59-65.

Tang, W. C., Lu, S. W., Tsai. X. M., Kao. C. Y., Lee. H.

H., 2007, Harmonic Parameters with HHT and

Wavelet Transform for Automatic Sleep Stages

Scoring. Proc. of World Academy of Science.

Engineering and Technology, 22, 414-417.

Yilmaz, A., Alkan, A., Asyali, M., 2007, Applications of

parametric spectral estimation methods on detection of

power system harmonics. Electric Power Systems

Research (2008), 78, 683-693.

Zoubek, L., Charbonnier, S., Lesecq, S., Buguet, A.,

Chapolot, F., 2007, Feature selection for sleep/wake

stages classification using data driven methods.

Biomedical Signal Processing and Control, 2, 171-

179.

Level 2

Level 3

Wake

Epoch

Sleep

REM

N2/N3

NREM

Level 1

Level 4

FEATURE EXTRACTION AND SELECTION FOR AUTOMATIC SLEEP STAGING USING EEG

133