EMD-BASED ANALYSIS OF RAT EEG DATA FOR

SLEEP STATE CLASSIFICATION

S¨uleyman Baykut

1,2

, Paulo Gonc¸alves

, Pierre-Herv´e Luppi

, Patrice Abry

Edmundo Pereira de Souza Neto

4,5

and Damien Gervasoni

Department of Electronics and Communications Engineering, Istanbul Technical University, Istanbul, Turkey

INRIA-RESO, LIP, Ecole Normale Sup´erieure de Lyon, Lyon, France

Physiopathologie des R´eseaux Neuronaux du Cycle Veille-Sommeil, UMR5167 CNRS, Universit´e Claude Bernard, Lyon, France

Laboratoire de Physique, UMR 5672 CNRS, Ecole Normale Sup´erieure de Lyon, Lyon, France

Service d’Anesth´esie R´eanimation, Hˆopital Pierre Wertheimer, Hospices Civils de Lyon, Bron, France

Keywords:

Automatic Sleep State Classiﬁcation, Empirical Mode Decomposition, Rat EEG.

Abstract:

In this paper Empirical Mode Decomposition (EMD)-based features from single-channel electroencephalo-

graphic (EEG) data are proposed for rat’s sleep state classiﬁcation. The classiﬁcation performances of the

EMD-based features and some classical power spectrum density (PSD)-based features are compared. Sup-

ported by experiments on real EEG data, we demonstrate that classiﬁcation performances can signiﬁcantly

improve, by simply substituting EMD to PSD in features extraction. This is in noticeably due to the natural

adaptivity of EMD which show more robust to subjects variability.

1 INTRODUCTION

EEG signals are widely used to study the basic as-

pects of the brain activities in human and/or in ani-

mals. EEG data show variations related to the differ-

ent brain states during sleep which make EEG an im-

portant tool in sleep studies (Robert et al., 1999; Vya-

zovskiy et al., 2002; Estrada et al., 2004; Hese et al.,

2001; Vivaldi and Bassi, 2006; Corsi-Cabrera et al.,

2001). Objective staging of the sleep states by visual

analysis of EEG data is a time consuming and rather

subjective process, as staging agreement between dif-

ferent experts is around 80 − 90% (Gervasoni et al.,

2004; Robert et al., 1999). In recent years, the re-

searches focused on deﬁning new methods, mostly

utilizing EEG data, in order to provide an automatic

and more objective staging of the brain states. How-

ever, due to the complexity and to the subject variabil-

ity of EEG signals, automatic sleep classiﬁcation still

remains a challenging issue.

Automatic sleep state classiﬁcation based on EEG

data has been performed by a large number of tech-

niques relying on time, frequency or wavelet do-

main features, such as EEG amplitudes, zero crossing

counting, harmonic analysis, Hjorth parameters, spe-

ciﬁc bandwidths’ energies (δ, θ, σ), bispectra (Ning

et al., 1990), to cite but a few.

In most studies dealing with automatic quantita-

tive classiﬁcation of sleep states, frequency domain

features seem to be preferred. In particular, Power

Spectrum Density (PSD)-based classiﬁcations rely on

the energy carried out in different frequency inter-

vals, even if a unanimous consensus about the deﬁ-

nition of sensitive bandwidths is still missing (Robert

et al., 1999; Corsi-Cabrera et al., 2001). It is also no-

ticed that mostrepresentativeEEG activities may vary

between sleep states and between different subjects

(Corsi-Cabrera et al., 2001). Moreover, to be fully

relevant, PSD-based methods require the data to be

stationary over the analyzedepoch. This is anotherse-

vere limitation when dealing with EEG signals which

by nature are non-stationary as it is the case for most

biomedical signals. For all these reasons, a method

that is adaptive (data-driven), less sensitive to subject

heterogeneity and robust to non-stationary data is cer-

tainly worth investigating.

Empirical Mode Decomposition (EMD) is known

to be a fully data-driven technique which automati-

cally extracts meaningful oscillating components (in-

trinsic modes) that underlie the signal (Huang et al.,

1998; Rilling et al., 2003). Furthermore, as it acts

locally in time, EMD does not require the analyzed

115

Baykut S., Gonçalves P., Luppi P., Abry P., Pereira de Souza Neto E. and Gervasoni D. (2009).

EMD-BASED ANALYSIS OF RAT EEG DATA FOR SLEEP STATE CLASSIFICATION.

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing, pages 115-123

DOI: 10.5220/0001546501150123

 SciTePress

signal to be stationary. The main goal of this study

is then to underline how the inherent advantages of

EMD can sensibly strengthen the features’ extraction

and improve the sleep state classiﬁcation.

The article is organized as follows. Data and

methods are described in Sections 2 and 3, respec-

tively. Section 4 presents and compares the classi-

ﬁcation performances of both PSD and EMD-based

approaches. We conclude and discuss possible exten-

sions of this work in Section 5.

2 MATERIALS

In order to reduce the number of experimented

animals, we advisedly re-used for this study a data

set originally obtained from previous experiments.

In these experiments, adult male Sprague Dawley

rats (250-300 g) were chronically implanted for

polygraphic recordings under general anesthesia

(Ketamine / Xylazine, 90 mg/kg and 5mg/kg re-

spectively). Brieﬂy, three stainless steel EEG screws

were inserted into the skull over the frontal, parietal

and occipital cortices, with a reference electrode

placed over the cerebellum. Two stainless wires were

inserted into the neck muscles. All electrodes were

linked to a connector cemented to the skull with

dental acrylic. Post-operative analgesia was ensured

by oral administration of carprofen (Rimadyl



, 0.5

mg / 100 g / 24 h). After ten days of recovery from

the surgical procedure, the rats were individually

habituated to the recording chamber and cable under

a 12 hour light/dark cycle (lights on at 7:00 A.M.)

with water and food ad libitum. All procedures

were in accordance with the National Institute of

Health guidelines for animal care and were ap-

proved by our local institutional ethics committee

for animal experimentation. From each rat, three

referential EEG and one bipolar EMG signals were

ampliﬁed and ﬁltered (bandwidth 1-250 Hz) using a

preampliﬁer headstage and a multichannel ampliﬁer

(MCP Plus, Alpha-Omega Engineering, Nazareth,

Israel), and sampled at 512 Hz using an analog to

digital converter (Cambridge Electronic Design,

Cambridge, UK). In this study, the data set from 6

rats are analyzed. Data ﬁles were manually scored by

10-seconds epochs using EEG and EMD magnitude

criteria deﬁned in the four distinct papers (Gervasoni

et al., 2004), (Gottesmann et al., 1976), (Timo-Iaria

et al., 1970), (Winson, 1974). Although the epochs

were scored by two scorers with an inter-rater agree-

ment of 85%, here the result from only one scorer

was used. Blind repeated scoring was not performed

in this study. Three sleep states are identiﬁed;

0 1 2 3 4 5 6 7 8 9 10

−4

−2

a) AW

0 1 2 3 4 5 6 7 8 9 10

−4

−2

b) SWS

Amplitude

0 1 2 3 4 5 6 7 8 9 10

−4

−2

c) PS

time − [sec]

Figure 1: Rat EEG Signals at a) Awake, b) Slow Wave

Sleep, c) Paradoxical Sleep.

1. Awake (AW): AW epochs are characterized by

low amplitude EEGs with high θ (5-9 Hz) and γ

(30-55 Hz) power density and a high amplitude

EMG with phasic bursts.

2. Slow Wave Sleep (SWS): SWS epochs are iden-

tiﬁed by high amplitude EEG with a high δ (1-4

Hz) and spindles (10-14 Hz) power density and a

low amplitude EMG.

3. Paradoxical Sleep (PS): PS (Rapid Eye Move-

ment) was identiﬁed by a low amplitude EEG

with a predominant θ rhythm (5-9 Hz) and a con-

comitant low amplitude EMG reﬂecting the typi-

cal muscular atonia.

In Figure 1 a-c, examples of 10-seconds rat EEG sig-

nals from frontal cortex derivation are plotted for AW,

SWS and PS states, respectively.

All the segments with saturated sample points are

discarded from the analysis. In the data set, only the

three main sleep states are identiﬁed. However, it

is known that, at the sleep cycles, middle transition

sates are also present between main sleep states (Ger-

vasoni et al., 2004). These transition states may cre-

ate ambiguity during the search for best representative

and discriminative features. Therefore, the transition

segments that may correspond to possible transition

states are also excluded from the data base. The num-

bers of remaining segments after data cleaning are de-

scribed in Table 1 for all states and from each of the

six rats.

Table 1: The number of segments at each sleep states for six

rats.

Rat Time \ segments # AW # SWS # PS

1 208 min \ 1253 863 198 192

2 91 min \ 547 200 135 212

3 115 min \ 689 352 188 149

4 241 min \ 1446 930 467 49

5 194 min \ 1162 766 232 164

6 310 min \ 1860 784 869 207

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

116

Table 2: Spectral Energy Bands.

Band Bandwidth (Hz)

0.5 - 2.5

2.5 - 4.0

4.0 - 6.0

6.0 - 8.0

α 8.0 - 12.0

12.0 - 20.0

20.0 - 45.0

3 METHODS

This section starts presenting two classical ap-

proaches based on the estimated power spectrum den-

sity of EEG time series. After describing the PSD-

based techniques, new feature sets based on EMD

analysis are proposed. Each data segment is normal-

ized before PSD or EMD analyses in order to have

unit energy within each 10-seconds segment.

3.1 PSD-based Features

3.1.1 Relative Bandwidths’ Energies

Several studies have shown relevance in parsing EEG

signals into different frequency bands. For instance,

(Goeller and Sinton, 1989) suggests to consider four

bands 1-3 Hz, 4-6 Hz, 6-10 Hz and 11-25 Hz; accord-

ing to (Ruigt et al., 1989), seven bands are relevant:

1-3 Hz, 3-6 Hz, 6-9 Hz, 9.5-20 Hz, 20-45 Hz, 49-51

Hz and 1-45 Hz; as for (Neckelmann et al., 1994),

only four bands are necessary:0.5-5 Hz, 6-9 Hz, 11-

16Hz, 20-60.5 Hz. We chose in the present study to

divide the frequency axis into seven different energy

bands as proposed in (Estrada et al., 2004; Hese et al.,

2001) and reported in Table 2.

The normalized PSD is estimated for each 10 sec-

onds EEG segment, and then partially integrated over

these seven frequency intervals. The resulting ener-

gies serve as input features of the sleep state classiﬁ-

cation.

3.1.2 PSD-based Spectral Energy Ratios

In the second PSD-based approachdeveloped in (Ger-

vasoni et al., 2004), two speciﬁc spectral energy ratios

are deﬁned and used as input features of a supervised

classiﬁer. In (Gervasoni et al., 2004), Local Field

Potential (LFP) data is analyzed in order to exam-

ine sleep state transitions. Four frequency bands are

deﬁned: 0.5-20Hz, 0.5-55H, 0.5-4.5Hz, 0.5-9Hz and

then the spectral amplitude ratios, r

and r

, are ob-

tained by dividing the integrated spectral amplitudes

over deﬁned frequency bands:

20Hz

0.5Hz

P( f)d f

55Hz

0.5Hz

P( f)d f

4.5Hz

0.5Hz

P( f)d f

9Hz

0.5Hz

P( f)d f

(1)

where P( f) denotes the spectral amplitudes over the

frequency f. These ratios are heuristically chosen af-

ter a thorough search in order to obtain the best sepa-

ration of states. Numerators of the ratios are included

in the denominators to obtain more symmetrical dis-

tributions. The feature space for classiﬁcation is con-

structed by two variables (r

, r

) calculated from each

segment.

3.2 EMD-based Features

EMD is an entirely data-driven (adaptive)method that

iteratively decomposes the analyzed signal into a set

of componentscalled Intrinsic ModeFunctions (IMF)

(Huang et al., 1998). In contrast with the Fourier or

the wavelet analyses, EMD adaptively extracts the in-

trinsic components that compose the signal without

necessitating to choose any a priori ﬁxed basis. The

core of the algorithm is the so-called “sifting proce-

dure” which, locally in time, isolates the fastest os-

cillation in the signal. The resulting (possibly non-

stationary) component, referred to as the ﬁrst intrinsic

mode function (IMF), is an amplitude and frequency

modulated waveform that corresponds to the signal

details at the ﬁnest time scales. This IMF is subtracted

from the original signal, and the same adaptive proce-

dure is applied to the remainder (low-pass approxi-

mation) to identify the second IMF, and recursively

for the next ones.

EMD has been successfully used in many biolog-

ical signal processing applications (see e.g. (Ziqiang

and Puthusserypady, 2007; Torres et al., 2007; Shara-

baty et al., 2006)), and we believe that EMD-based

features are also promising candidates for a more ro-

bust sleep state classiﬁcation.

EMD is applied to each of the EEG signal ex-

amples shown in Figure 1 and the ﬁrst 8 IMFs of

each signal are displayed in Figure 2-a. It was ob-

served that a N points time series generally decom-

poses into log

(N) IMFs (Huang et al., 1998; Rilling

et al., 2003). In our case, we most often obtained be-

tween 10 and 12 IMFs but only the ﬁrst 8 are sys-

tematically considered as the frequencycontent of the

remainders lay below the physiologically meaningful

bandwidth (< 0.5Hz). As it can be seen in the ﬁg-

ure, the IMFs reproduce the non-stationarities of the

signal at different characteristic time scales.

EMD-BASED ANALYSIS OF RAT EEG DATA FOR SLEEP STATE CLASSIFICATION

117

1000 2000 3000 4000 5000

−0.5

0.5

# 1

Awake

1000 2000 3000 4000 5000

−0.5

0.5

Slow Wave Sleep

1000 2000 3000 4000 5000

−0.5

0.5

Paradoxial Sleep

1000 2000 3000 4000 5000

−1

# 2

1000 2000 3000 4000 5000

−1

1000 2000 3000 4000 5000

−1

1000 2000 3000 4000 5000

−1

# 3

1000 2000 3000 4000 5000

−1

1000 2000 3000 4000 5000

−1

1000 2000 3000 4000 5000

−1

# 4

1000 2000 3000 4000 5000

−1

1000 2000 3000 4000 5000

−1

1000 2000 3000 4000 5000

−1

# 5

1000 2000 3000 4000 5000

−1

1000 2000 3000 4000 5000

−1

1000 2000 3000 4000 5000

−0.2

0.2

# 6

1000 2000 3000 4000 5000

−0.2

0.2

1000 2000 3000 4000 5000

−0.2

0.2

1000 2000 3000 4000 5000

−0.1

0.1

# 7

1000 2000 3000 4000 5000

−0.1

0.1

1000 2000 3000 4000 5000

−0.1

0.1

1000 2000 3000 4000 5000

−0.05

0.05

0.1

# 8

1000 2000 3000 4000 5000

−0.05

0.05

0.1

1000 2000 3000 4000 5000

−0.05

0.05

0.1

0.5 2 5 10 25 50 100

0.02

0.04

0.06

0.08

0.1

log(f)

Amplitude

0.5 2 5 10 25 50 100

0.02

0.04

0.06

0.08

0.1

SWS

log(f)

Amplitude

0.5 2 5 10 25 50 100

0.02

0.04

0.06

0.08

0.1

log(f)

Amplitude

IMF−1

IMF−2

IMF−3

IMF−4

IMF−5

IMF−6

IMF−7

IMF−8

IMF−1

IMF−2

IMF−3

IMF−4

IMF−5

IMF−6

IMF−7

IMF−8

IMF−1

IMF−2

IMF−3

IMF−4

IMF−5

IMF−6

IMF−7

IMF−8

8 7 6 5 4 3 2 1

0.5

log(f)

IMF #

8 7 6 5 4 3 2 1

0.5

SWS

log(f)

IMF #

8 7 6 5 4 3 2 1

0.5

log(f)

IMF #

Figure 2: a) The ﬁrst 8 IMFs of EEG Signals for AW (the 1

column), for SWS (the 2

column) and for PS (the 3

column)

states; b) Average PSD of IMFs at the states AW, SWS and PS (from left to right respectively); c) The logarithm of mean

values and the standard deviations of the peak frequencies of IMF spectra (EEG data segments are taken from rat-1).

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

118

Note that, for broad band signals such as EEG,

EMD behaves like a dyadic ﬁlter bank (Flandrin et al.,

2003). In Figure 2-b, average PSDs of IMFs obtained

from EEG segments (at AW, SWS and PS states) are

given. It can be observed that the frequency struc-

tures of the automatically generated IMFs are similar

to that of artiﬁcially pre-deﬁned EEG bands. For ex-

ample, the frequency content of the ﬁrst IMF approx-

imately corresponds to the γ band (>30Hz), whereas

the second IMF overlaps with the β band (13 − 30

Hz). The third and the fourth IMFs also coincide with

the α (8− 13 Hz) and θ (3.5− 8 Hz) bands. A still

open question though, is to determine to which extent

these adaptively identiﬁed bandwidths correspond to

physical phenomena. Note that the frequency regions

that are automatically extracted by EMD are variable,

i.e. they are changing adaptively with the analyzed

signal. This can be seen from Figure 2-b that com-

pares the IMF spectra corresponding to the differ-

ent states. Moreover, the frequency peak distribution

(mean and variance) of each IMF spectrum, displayed

as box plots in Figure 2-c, show that characteristic

modes of a given state can also vary along time.

3.2.1 EMD-based Band Energies

As it is seen in Figure 2-b, the energies at some scales

are noticeably different for different states which can

be convenient for classiﬁcation. Therefore, energies

of eight IMFs are individually considered as features

for sleep staging. This approach is similar to the

PSD-based relative frequency band energy features

given in Section 3.1.1. However, instead of deﬁning

ﬁxed bands beforehand, the bands are automatically

selected by the method.

3.2.2 EMD-based Spectral Energy Ratios

By considering the same frequency bands deﬁned in

Section 3.1.2 the EMD-based spectral ratios can ap-

proximately mimic the PSD-based spectral ratios. It

is approximately identiﬁed which IMF lie in which

bandwidths considered in the energy ratios r

and r

deﬁned in (1). The EMD based energy ratios are then

constructed as below:

∑

i∈(2,..,7)

IMF

∑

i∈(1,..,7)

IMF

∑

i∈(5,..,7)

IMF

∑

i∈(4,..,7)

IMF

(2)

Here P

IMF

is the energy of the i

IMF.

3.2.3 EMD-based New Spectral Ratios

EMD providelimited number of naturally dividedfre-

quency bands which make possible to search for ad-

ditional features. In Figure 3, the energy spectra of

IMFs are plotted for three sleep states. All six rats’

data are combined and the mean and the standard de-

viations of the IMF energies are calculated. This ﬁg-

ure is utilized to search for new energy ratios that may

provide better separation of the classes.

8 7 6 5 4 3 2 1

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

IMF #

Mean IMF Energies +− σ

IMF Enegry Spectrum

SWS

Figure 3: IMF Energy Spectra (the mean and the standard

deviation of the IMF energies) of three sleep states.

We deﬁne two new ratios. The ﬁrst ratio is deﬁned

for separation of AW and SWS states. It is slightly

different than the ratio in (2) (6

and the 7

IMF en-

ergies are removed from the ratio since they are not

signiﬁcantly different for AW and SWS states as it is

seen in Figure 3).

The second ratio is deﬁned for the separation of

AW and PS states. It is observed in Figure 3 that the

energies at the ﬁrst two IMFs obtained from PS state

are higher than the energies of IMFs obtained from

AW state. In contrast, the energies of IMFs from AW

at last four scales (5, 6, 7, 8) are higher than the ener-

gies of IMFs from PS state. Considering these obser-

vations, the second ratio is deﬁned as the total energy

of the ﬁrst two IMFs over the total energy of the last

four IMFs. Similar to the previously deﬁned ratios,

the denominator includes also the numerator compo-

nents:

∑

i∈(2,..,5)

IMF

∑

i∈(1,..,5)

IMF

∑

i∈(1,2)

IMF

∑

i∈(1,2,5,6,7,8)

IMF

(3)

The two PSD-based and the threeEMD-based fea-

ture sets deﬁned in the previous Sections are used as

input features of a supervised classiﬁer of EEG data.

Results are compared in the next section.

4 STATISTICAL ANALYSIS AND

RESULTS

A simple supervised classiﬁer is used by ﬁtting a mul-

tivariate normal density to each class. Several classi-

EMD-BASED ANALYSIS OF RAT EEG DATA FOR SLEEP STATE CLASSIFICATION

119

ﬁcation parameters are provided for each feature set

as a measure of classiﬁcation performances:

• Sensitivity (Sns.) and Speciﬁcity (Spc.): Sensi-

tivity is the proportion of the correctly identiﬁed

(true positives) segments that actually belong to

the class and speciﬁcity is the proportion of the

correctly identiﬁed (true negatives) segments that

actually do not belong to the class.

• Omission Error (ε

) and Commission Error (ε

is the proportion of the omitted segments that

actually belong to the class (ε

= 1 - Sns.). ε

is the

proportion of the segments classiﬁed as the corre-

sponding class where they actually do not belong

to the class (ε

= 1 - Spc.).

• Overall Accuracy (Acc.): Accuracy is the propor-

tion of the truly classiﬁed segments in the whole

population.

• Kohen’s Kappa (κ): κ is a measure of agree-

ment between the manual and automatic staging.

1 > κ > 0.81 indicates almost perfect agreement,

whereas 0.8 > κ > 0.61 and 0.6 > κ > 0.41 indi-

cate substantial and moderate agreements respec-

tively. This parameter is different from the overall

agreement in the sense that it removes the effects

of the agreements occurred by chance. κ is calcu-

lated as: κ =

−P

1−P

where P

is the overall agree-

ment between manual and automatic staging, and

is the expected agreement by chance (Cohen,

1960).

• ConcordanceTable (Confusion Matrix): The frac-

tion of the data which was classiﬁed as each of

the existing classes are given in concordance ta-

bles. In these tables, each row represents the ac-

tual class whereas each column representsthe pre-

dicted class.

We hold two different scenarios of classiﬁcation ex-

periment.

• Self Classiﬁcation: In the ﬁrst group experiment,

all the data from the six rats are gathered and a

5-folds cross validation is used. Since the training

and the test data are selected from the same data

pool, this experiment is called self-classiﬁcation

throughout the paper.

• Cross Classiﬁcation: In the second group exper-

iment, we train the classiﬁer with mixed data

coming from only ﬁve rats, and test the algo-

rithm with the remaining sixth rat’s EEG. This

procedure is repeated for all possible combina-

tions of the rats and only the mean classiﬁcation

rates are provided. We refer to this experiment as

cross-classiﬁcation. The aim of this experiment

is to assess the robustness of the features to inter-

individual variability.

4.1 Self-classiﬁcation Experiments

4.1.1 PSD-based Band Energies vs. EMD-based

Band Energies

Self-classiﬁcation results (mean ± standard devia-

tion of each performance parameters) obtained by the

PSD-based and EMD-based band energy features are

given in Tables 3 and 4. It is observed in the con-

cordance tables that the separation of SWS class from

the other two classes is attains a high matching score

(with ∼ 96− 98%) by both PSD and EMD based fea-

tures. On the other hand, separation of AW and PS

states is more arduous. Yet, the energies computed

over the adaptively EMD-selected bandwidths pro-

vide relatively better results. Indeed, the omission

error of AW state decreases of about 7% and the com-

mission error of PS state gains 18%. As we can see,

simply replacing the pre-ﬁxed bands with the ones

that are automatically identiﬁed by EMD provides a

3% increase on the overall accuracy and a 0.04 me-

lioration of κ index. This shows that even within the

same rat’s EEG, the frequency content relative to a

given state, can vary from one time segment to the

other (as it can be seen in Figure 2-c), and that auto-

matic band selection is an encouraging alternative to

handle this non-stationarity.

4.1.2 PSD-based Energy Ratios vs. EMD-based

Energy Ratios

Self classiﬁcation results for PSD-based spectral ra-

tios and for their EMD-based counterpart are given in

Table 5 and 6, respectively. Similarly to the band en-

ergy features, SWS class is easily distinguished from

the other two classes (∼ 96−97%agreement) by both

PSD and EMD-based ratios. However, discrimination

of AW and PS classes is relatively worse. For PSD-

based ratios, 23% of AW segments are classiﬁed as

PS and 15% of PS segments are classiﬁed as AW. By

using EMD-based ratios, these percentages lower to

18% and 12% respectively. Overall accuracy also gets

better, with a 3% upgrade, whereas κ index increases

of about +0.04.

As it is explained in Section 3.2.3, in order to have

better separation between AW and PS classes, new

EMD-based ratios were deﬁned. The classiﬁcation

results are provided in Table 7. Compared to the PSD-

based ratios’ results of Table 5, EMD-based new ra-

tios pull the omission for AW class from 24% down to

17%. Omission of PS class also decreases from 16%

to 11%. Overall accuracy improves by 4.6% and the

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

120

κ gain is 0.07. For similar reasons to the ones evoked

in the preceding paragraph, it is very likely that the

adaptivity of EMD turns the energy ratios more ro-

bust to non-stationarities.

4.2 Cross-classiﬁcation Experiments

The aim of this experiment is to investigate the ro-

bustness of the feature sets to the inter-individualvari-

ability. The results obtained for the cross classiﬁ-

cation experiment are presented in Table 8. As for

the self classiﬁcation, EMD features globally provide

better classiﬁcation performances. Again the differ-

ences stem from a better separation between AW and

PS states. For example, compared to PSD-based en-

ergy ratios, EMD-based new energy ratios increases

the sensitivity for AW and PS classes by 9% and 7%,

respectively. The speciﬁcities also improve (4% and

13%), as well as the overall accuracy (about 5.5%)

and the κ index which raises by +0.09. Improvements

of the same order of magnitude can be observed in the

same table for frequency bands features. Altogether,

these results demonstrate that EMD-based features,

through their adapted bandwidth selection, ﬁt better

the individual characteristics. Indeed, IMFs convey

a spectral information (notably the frequency band-

width) that can signiﬁcantly vary from one individual

to the other. Then, as the ratios deﬁned in expres-

sions (3) solely imply IMFs indices, they can natu-

rally adapt to the spectral speciﬁcities of each indi-

vidual.

For better visual comparison, overall accuracies

and κ values obtained with PSD and EMD-based fea-

ture sets are bar-plotted in Figures 4 and 5.

Another interesting outcome of this experiment is

that in cross classiﬁcation EMD-based energy ratios

outperform the EMD-based energy bands whereas it

is the opposite for self-classiﬁcation. The overall ac-

curacy and the κ values of EMD-based ratios are sen-

sibly the same for self and cross classiﬁcations. With

EMD-based energy bands, the performancedecreases

for the cross classiﬁcation. This can be explained as

follows. When test and training data are selected from

different subjects, energies at some scales (mostly

coarse scales corresponding to low frequency compo-

nents) may substantially vary affecting thus the clas-

siﬁcation performance. With energy ratios however,

since they combine the energies at different scales,

this local energy variability is less penalizing.

a) Accuracy

Accuracy %

PSD−Based EMD−Based

0.68

0.7

0.72

0.74

0.76

0.78

0.8

0.82

0.84

b) Cohen’s κ

PSD

Bands

EMD

Bands

PSD

Ratio

EMD

Ratio

EMD

New

Ratio

PSD

Bands

EMD

Bands

PSD

Ratio

EMD

Ratio

EMD

New

Ratio

Figure 4: Final comparison of the classiﬁcation perfor-

mances (First group classiﬁcation).

a) Accuracy

Accuracy %

PSD−Based EMD−Based

0.68

0.7

0.72

0.74

0.76

0.78

0.8

0.82

0.84

b) Cohen’s κ

PSD

Bands

EMD

Bands

PSD

Ratio

EMD

Ratio

EMD

New

Ratio

PSD

Bands

EMD

Bands

PSD

Ratio

EMD

Ratio

EMD

New

Ratio

Figure 5: Final comparison of the classiﬁcation perfor-

mances (Second group classiﬁcation).

5 CONCLUSIONS AND FUTURE

WORKS

In this paper EMD is applied to rat EEG signals in

order to extract features for sleep state classiﬁcation.

The results obtained from EMD-based features are

compared to those obtained from PSD-based features,

and show that EMD-based techniques are particularly

adapted to analyze non-stationary signals, such EEGs

or any otherbiophysical signals. Since EMDis a data-

driven technique that naturally decomposes the data

into intrinsic components, it removes the necessity of

a hazardous ﬁxed band division. Adaptivity of EMD

also yields a more robust classiﬁcation with respect to

the inter-individual variability.

In addition, as the sifting process performs locally

in time, a on-line version of the EMD algorithm was

proposed in (Rilling et al., 2003). Then, we could

use this to adapt our EMD-based classiﬁer to contin-

uous time, with no prior segmentation of the signal

needed. This approach would permit sleep state mon-

itoring with real-time classiﬁcation and detection of

transition points.

EMD-BASED ANALYSIS OF RAT EEG DATA FOR SLEEP STATE CLASSIFICATION

121

Table 3: Self classiﬁcation performances after using PSD-based spectral band energy features.

AW SWS PS ε

AW 0.84±0.01 0.02±0.00 0.14±0.01 0.16±0.01 0.07±0.00

SWS 0.02±0.00 0.98±0.00 0.00±0.00 0.02±0.01 0.04±0.01

PS 0.20±0.01 0.01±0.00 0.79±0.01 0.21±0.02 0.42±0.01

Spc.=82.5%±0.4 Sns.=86.8%±0.4 Acc.=87.5%±0.4 κ=0.79±0.01

Table 4: Self classiﬁcation performances after using EMD-based spectral band energy features.

AW SWS PS ε

AW 0.91±0.01 0.03±0.00 0.06±0.01 0.09±0.01 0.08±0.01

SWS 0.04±0.01 0.96±0.01 0.00±0.00 0.04±0.01 0.06±0.01

PS 0.22±0.02 0.01±0.01 0.77±0.02 0.23±0.02 0.25±0.02

Spc.= 87.1%±0.7 Sns.=87.9%±0.7 Acc.=90.4%±0.4 κ=0.83±0.01

Table 5: Self classiﬁcation performances after using PSD-based spectral energy ratios.

AW SWS PS ε

AW 0.76±0.02 0.01±0.00 0.23±0.02 0.24±0.02 0.07±0.01

SWS 0.03±0.01 0.97±0.01 0.01±0.00 0.03±0.01 0.03±0.00

PS 0.16±0.02 0.00±0.01 0.84±0.02 0.16±0.03 0.53±0.02

Spc.=79.4%±0.6 Sns.=85.4%±0.8 Acc.=83.2%±0.9 κ=0.73±0.01

Table 6: Self classiﬁcation performances after using EMD-based spectral energy ratios.

AW SWS PS ε

AW 0.80±0.01 0.02±0.00 0.18±0.01 0.20±0.01 0.06±0.01

SWS 0.03±0.01 0.96±0.00 0.01±0.00 0.04±0.00 0.04±0.01

PS 0.12±0.02 0.01±0.00 0.88±0.02 0.12±0.02 0.46±0.02

Spc.=81.6%±0.9 Sns.=87.8%±1.0 Acc.=85.8%±0.9 κ=0.77±0.01

Table 7: Self classiﬁcation performances after using EMD-based new ratios.

AW SWS PS ε

AW 0.83±0.01 0.02±0.00 0.15±0.01 0.17±0.01 0.05±0.01

SWS 0.03±0.01 0.96±0.01 0.01±0.00 0.04±0.01 0.04±0.01

PS 0.11±0.02 0.01±0.00 0.89±0.02 0.11±0.02 0.41±0.02

Spc.=83.4%±0.7 Sns.=89.2%±0.9 Acc.=87.8%±0.6 κ=0.80±0.01

Table 8: Cross classiﬁcation results. The data from ﬁve rats are used for training and the remaining rat’s data is used for

testing. The procedure repeated for all rats and the mean values are provided.

AW SWS PS Mean Mean

Methods Spc. Sns. Spc. Sns. Spc. Sns. Spc. Sns. Acc. κ

PSD-Bands 89.7 79.2 92.7 96 56.7 74.8 79.7 83.3 84.0 0.72

EMD-Bands 91.5 80.2 87.1 98.4 69.5 78.7 82.7 85.8 85.5 0.75

PSD-Ratios 89.5 73.6 98.2 91.4 47.5 81.8 78.4 82.3 81.9 0.69

EMD-Ratios 93.0 79.5 96.7 95.7 54.7 87.1 81.5 87.4 85.7 0.76

EMD-New 93.4 82.2 96.6 95.3 60.0 88.1 83.3 88.6 87.4 0.78

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

122

ACKNOWLEDGEMENTS

S¨uleyman Baykut’s work is supported by “The

Scientiﬁc and Technological Research Council of

Turkey-The Department of Science Fellowships and

Grant Programmes (TUBITAK-BIDEB)” with the

programme # 2214.

REFERENCES

Cohen, J. (1960). A coeeﬁcient of agreement for nominal

scales. In Educational and Psychological Measure-

ment. Vol. 20, pp. 37-46.

Corsi-Cabrera, M., P´erez-Garci, E., Rio-Portilla, Y. D.,

Ugalde, E., and Guevara, M. A. (2001). EEG bands

during wakefulness, slow-wave, and paradoxical sleep

as a result of principal component analysis in the rat.

In Sleep. Vol. 24, No. 4.

Estrada, E., Nazeran, H., Nava, P., Behbehani, K., Burk,

J., and Lucas, E. (2004). EEG feature extraction for

classiﬁcation of sleep stages. In Proc. of the 26th An-

nual International Conference of the IEEE EMBS. San

Francisco, USA.

Flandrin, P., Rilling, G., and Goncalves, P. (2003). Empiri-

cal mode decomposition as a ﬁlter bank. In IEEE Sig.

Proc. Lett. Vol. 11, No. 2, Part 1, pp. 112- 114.

Gervasoni, D., Lin, S., Ribeiro, S., Soares, E., Pantoja, J.,

and Nicolelis, M. (2004). Global forebrain dynam-

ics predict rat behavioral states and their transitions.

In The Journal of Neuroscience. Vol. 24, No. 49, pp.

11137-11147.

Goeller, C. J. and Sinton, C. M. (1989). A microcomputer-

based sleep stage analyser. In Computer Methods and

Programs in Biomedicine. Vol. 29, No. 1, 316.

Gottesmann, C., Lacoste, G., Rodrigues, L., Kirkham, P.,

Arnaud, C., and Rallo, J. L. (1976). Method of auto-

matic analysis and quantiﬁcation of wakefulness-sleep

behavior in the rat. In Rev. Electroencephalogr. Neu-

rophysiol. Clin. Vol. 6, pp. 37-49.

Hese, P. V., Philips, W., Koninck, J. D., de Walle, R. V.,

and Lemahieu, I. (2001). Automatic detection of sleep

stages using the EEG. In Proc. of the 23rd Annual

International Conference of the IEEE EMBS. Istanbul,

Turkey.

Huang, N., Shen, Z., Long, S., Wu, M., Shih, H., Zheng,

Q., Yen, N. C., Tung, C. C., and Liu, H. (1998). The

empirical mode decomposition and hilbert spectrum

for nonlinear and nonstationary time series analysis.

In Proc. Roy. Soc. London A. Vol. 454, pp. 903-995.

Neckelmann, D., Olsen, O. E., Fagerland, S., and Ursin,

R. (1994). The reliability and functional validity of

visual and semi-automatic sleep:wake scoring in the

moll-wistar rat. In Sleep. Vol. 17, No. 2, pp. 12031.

Ning, T., Joseph, D., and Bronzino, D. (1990). Bispectral

analysis of the rat EEG during REM sleep. In Proc.

of the Annual International Conference of the IEEE

EMBS. Vol. 12, No. 5.

Rilling, G., Flandrin, P., and Goncalves, P. (2003). On em-

pirical mode decomposition and its algorithms. In

IEEE-EURASIP Workshop on Nonlinear Signal and

Image Processing NSIP-03. Grado.

Robert, C., Guilpin, C., and Limoge, A. (1999). Automated

sleep staging systems in rats. In Journal of Neuro-

science Methods. Vol. 88, No.2, pp. 111-122.

Ruigt, G. S. F., Proosdij, J. N. V., and Delft, A. M. L. V.

(1989). A large scale, high resolution, automated sys-

tem for rat sleep staging i. In Methodology and techni-

cal aspects. Electroencephalogr. Clin. Neurophysiol.

Vol. 73, pp. 52-63.

Sharabaty, H., Martin, J., Jammes, B., and Esteve, D.

(2006). Alpha and theta wave localisation using

hilbert-huang transform: Empirical study of the ac-

curacy. In Proc. of IEEE Information and Communi-

cation Technologies-ICTTA. Vol. 1, pp. 1159-1164.

Timo-Iaria, C., Negrao, N., Schmidek, W. R., Hoshino, K.,

de Menezes, C. E. L., and da Rocha, T. L. (1970).

Phases and states of sleep in the rat. In Physiol. Behav.

Vol. 5, pp. 1057-1062.

Torres, A., Fiz, J. A., Jan´e, R., Galdiz, J. B., Gea, J., and

Morera, J. (2007). Application of the empirical mode

decomposition method to the analysis of respiratory

mechanomyographic signals. In Proc. of the 29th

Annual International Conference of the IEEE EMBS.

Lyon, France.

Vivaldi, E. A. and Bassi, A. (2006). Frequency domain

analysis of sleep EEG for visualization and automated

state detection. In Proc. of the 28th Annual Interna-

tional Conference of the IEEE EMBS. New York City,

USA.

Vyazovskiy, V. V., Borb´ely, A. A., and Tobler, I. (2002).

Interhemispheric sleep EEG asymetry in the rat is en-

hanced by sleep deprivation. In J. Neurophysiology.

Vol. 88, pp. 2280-2286.

Winson, J. (1974). Patterns of hippocampal theta rhythm

in the freely moving rat. In Electroencephalogr. Clin.

Neurophysiol. Vol. 36, pp. 291-301.

Ziqiang, Z. and Puthusserypady, S. (2007). Analysis of

schizophrenic EEG synchrony using empirical mode

decomposition. In Proc. of 15th Int. Conference on

Digital Signal Processing. pp. 131-134, Cardiff, Eng-

land.

EMD-BASED ANALYSIS OF RAT EEG DATA FOR SLEEP STATE CLASSIFICATION

123