A Channel Selection Method for EEG Classiﬁcation based on

Exponentially Damped Sinusoidal Model

and Stochastic Relevance Analysis

Leonardo Duque Mu

noz

, Carlos Guerrero-Mosquera

and German Castellanos-Dominguez

Grupo de Autom

atica y Electr

onica, Laboratorio MIRP, Instituto Tecnol

ogico Metropolitano, Medell

ın, Colombia

Departamento de Ingenier

ıa Electrica Electr

onica y Computaci

on, Universidad Nacional de Colombia,

Medell

ın, Colombia

Keywords:

Rhythms Decomposition, Seizure Detection, Feature Extractions, EEG Classiﬁcation.

Abstract:

This work introduces a new methodology to select EEG channels related to epileptic seizures by electroen-

cephalogram (EEG) rhythms extraction. Rhythms extraction is an alternative to extract useful information

from speciﬁc band frequencies, analyze changes in the EEG signals, and detect brain abnormalities. In this

approach, the EEG signals are modeled by Exponentially Damped Sinusoidal model (EDS) and the EEG

rhythms extraction is based on Stochastic Relevance Analysis (SRA). Achieve results show that EDS model

combined with a stochastic relevance measure is a proper alternative for EEG classiﬁcation of epileptic signals

and also could be used for EEG channel selection with seizure activity. The effectiveness of this approach is

compared in each experiment with other well known method for feature extraction called as Rhythmic Com-

ponent Extraction (RCE). This comparison was done based on the performance of the k-NN classiﬁers and

the channels selected were validated by visual inspection and topographic scalp map. The study uses real and

multi-channel EEG data and all the experiments have been supervised by an expert neurologist. We conclude

that the proposed scheme is a suitable approach for automatic seizure detection at a moderate computational

cost, also opening the possibility of formulating new criteria to select, classify or analyze abnormal EEGs

channels.

1 INTRODUCTION

Electroencephalographic signals (EEG) contain use-

ful information about the condition of the brain state

and their analysis is important to extract hidden infor-

mation or features that the human cannot directly de-

tect by visual inspection. Moreover, EEG signals with

epilepsy change continuously and to process every

EEG channel requires a long period of time, more vi-

sual analysis, high storage databases, and more com-

putational cost. To improve the efﬁciency in EEG

processing it is necessary to deal with multi-channel

processing and highlight the most pertinent features

of the signal. This allows to identify and extract EEG

rhythms faster, and also to see which parts of the brain

are the most affected for some abnormality.

Several authors have shown that epileptic seizures

could be decomposed into one or more components.

For example, the typical pattern or ictal rhythm in

mesial temporal lobe epilepsy appears as a high volt-

age blunt or sharp 5 − 7 Hz theta rhythm over one te-

mporal region (Nam et al., 2002). Other authors show

a principal epileptic track in band frequency range

between 2 − 7 Hz (Guerrero-Mosquera et al., 2010).

Therefore, EEG rhythms extraction could be used in

detecting brain abnormalities such as epilepsy dis-

ease.

There are different approaches propose in the liter-

ature for extracting EEG rhythms, such as wavelet de-

composition (Zalay et al., 2009), Independent Com-

ponent Analysis (ICA) (Sri and Rajapakse, 2008),

adaptive schema (Veluvolu et al., 2012), multi-

dimensional decomposition (Orekhova et al., 2011)

and frequency dominant characterization (Lodder and

Putten, 2011). Most of these studies are only fo-

cused on ﬁnding the onset and duration of the rhythm,

present many parameters to validate, show a depen-

dency with frequency resolution, but do not reﬂect the

true peak amplitude of the posterior dominant rhythm.

Other well know technique is the Rhythmic Com-

ponent Extraction (RCE), which combines multi-

channel signals with weights that are optimally sought

284

Duque Muñoz L., Guerrero-Mosquera C. and Castellanos-Dominguez G..

A Channel Selection Method for EEG Classiﬁcation based on Exponentially Damped Sinusoidal Model and Stochastic Relevance Analysis.

DOI: 10.5220/0004196802840289

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2013), pages 284-289

ISBN: 978-989-8565-36-5

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

for such that the extracted component maximally con-

tains the power in the frequency range of interest and

suppresses that in unnecessary frequencies

. Other

methods proposed at the state-of-the-art need to vali-

date many parameters or demand computational cost

in training (Peters et al., 2001; Yang et al., 2012;

Duun-Henriksen et al., 2012).

In this paper, we introduce a new method for EEG

channel selection based on Exponentially Damped

Sinusoidal model (EDS) and Stochastic Relevance

Analysis (SRA). Our approach distinguishes variables

that represent effectively a “hidden” phenomena ac-

cording to stochastic variability measure, that is used

as relevance function (Sepulveda-Cano et al., 2011),

and detects the EEG channels with more seizure ac-

tivity. The effectiveness of our approach is compared

with RCE method both for EEG segment classiﬁca-

tion problems and EEG channel selection. The classi-

ﬁer used is the well known k-nearest neighbor (k-nn)

algorithm.

This paper is organized as follows: Section 2 in-

troduces the EDS model and the Stochastic Rele-

vance Analysis (SRA). Section 3 describes the EEG

databases and the experiments setup, and ﬁnally in

Section 4 and Section 5, the main results are discussed

and the principal conclusions with further work are

presented.

2 PROCESSING METHODS

The proposed method is based on two main steps:

(1) EEG is modeled by Exponentially Damped Si-

nusoidal model (EDS), and (2) EEG rhythms are

extracted by Stochastic Relevance Analysis (SRA).

Both methods are brieﬂy explained as follows.

2.1 Exponentially Damped Sinusoidal

Model (EDS)

EDS model represents an EEG signal as a ﬁnite sum

of p discrete-time exponentially damped complex si-

nusoids:

x(n) =

∑

i=1

exp( j

)exp((−d

+ j2

)n∆t) +

(n)

n = 0, 1,...,N − 1

where n∆t is the time lapse between the origin and the

sample x(n), ∆t is the sampling interval, and

(n) is

Gaussian white noise. The parameters of this model

are the amplitudes a

, the phases

, the damping

More details of RCE algorithm in (Tanaka and Saito,

2008)

factors d

, and the frequencies f

. Since EDS are

a set of basis functions, they can be used to model

any arbitrary signal sufﬁciently closely, assuming that

the model order is high enough. Present study ad-

dresses the problem of estimating the model param-

eters when the EEG signal is embedded in noise by

using the subspace–based exponential data ﬁtting pro-

posed in (De Clercq et al., 2005), where EEG signal

is stacked in Hankel data matrix. Then, the model pa-

rameters needed in Eq. (1) are estimated by Singular

Value Decomposition (SVD). More detail of EDS pa-

rameters estimation in (De Clercq et al., 2005).

2.2 Stochastic Relevance Analysis

(SRA)

Relevance analysis distinguishes variables that rep-

resente effectively the subjacent physiological phe-

nomena according to some evaluation measure. Such

representative variables are named relevant features,

whereas the evaluation measure is known as relevance

measure. Variable selection tries to reject those vari-

ables whose contribution to representation target is

none or negligible (irrelevant features), as well as

those that have repeated information (redundant fea-

tures). Thus, the ﬁrst objective concerning the vari-

able selection stage is to deﬁne the concept of rele-

vance (Sepulveda-Cano et al., 2011). Let the set of

objects S

= {S

,k = 1,...,M} with M observations

described by a set of features s

i j

. In addition, each

sample is associated with one and only one element of

the set of class labels c

c =

{

(k)

∈ N : k = 1,··· ,K

}

where K is the number of classes to be considered.

Then, given S

, and for any feature s

i j

, the relevance

function

is deﬁned as:

: R

1×T

−→ R

i j

) 7→

i j

) ∈ R (1)

where the relevant function

satisﬁes the following

properties (Sepulveda-Cano et al., 2011):

– Non-negativity,

i j

) ≥ 0, for all i.

– Nullity, the function

i j

) is null if feature s

i j

has not relevance at all.

– Non-redundancy, if s

′

i j

, where

the real-valued

̸= 0, y

is some noyse

with zero mean and unit variance, then,



′

i j

) −

i j

)



→ 0

It is assumed that higher weights are associated with

the most relevant features. This work considers the

following unsupervised measure of relevance:

AChannelSelectionMethodforEEGClassificationbasedonExponentiallyDampedSinusoidalModelandStochastic

RelevanceAnalysis

285

c. Stochastic Variability: the time-varying relevance

measure is evaluated:

;

) = [

(1) ···

(

) ···

(pT )]

⊤

, (2)

where

(

) = E {|

(

)|}, {

: j = 1,...,q} is

the set of most relevant eigenvalues of the matrix

, and the scalar v

(

) is the respective element

in the instant

, and

= 1,..., pT indexes each

of the relevance values calculated for the entire

set of time-varying data. To determine the rele-

vance related to each of the stochastic variables,

the Eq.(2) can be arranged to the relevance

matrix [

;t)···

;t)]

⊤

where each row

;t) = [

(( f − 1)T +

1)...

(t)...

( f T )] ∈ R

T ×1

shows the contribu-

tion of the s

stochastic feature along ﬁxed time

moments.

3 MATERIAL AND SETTING

This work uses two EEGs databases: one of them

(labeled as DB1), described in detail in (Guerrero-

Mosquera et al., 2010), consists in 6 adult epilep-

tic patients obtained in a restful wakefulness stage

and recorded at the Clinica Universitaria de Navarra,

Department of Neurophysiology (Pamplona, Spain).

All of them contained focal epileptiform activity,

according to experienced neurologists. We used 6

EEG recordings of 24 min length taken from 23-

th, 24-th and 25-th channels using the 10- 20 Inter-

national System of Electrode Placement with addi-

tional anterotemporal electrodes T1/T2. In practice,

raw EEG were ﬁltered by a digital low-pass ﬁlter

with cut-off frequency of 20 Hz, sampled frequency

of 200 Hz and ocular artifacts rejected by Indepen-

dent Component Analysis (Guerrero-Mosquera and

Navia-Vazquez, 2012).

The other database (labeled as DB2) is a collec-

tion recorded at the Instituto de Epilepsia y Parkin-

son del Eje Cafetero (Pereira,Colombia). Each data

set from 35 patients contains 160 recorded scalp EEG

signals from 23-th, 24-th and 25-th channels corre-

sponding to electrodes placed on the head according

to the International 10-20 System of Electrode Place-

ment Standard. Set A holds 80 recordings labeled as

normal (seizure–free), whereas set E holds 80 record-

ings labeled as having epilepsy (epileptiform activity

by neurologist). Recordings have been sampled at a

frequency of 256 Hz, with 12 bits resolution and a 2-

min duration. All patients underwent clinical exami-

nation by neurologist. The data had been acquired in a

non–regulated conditions, and the noised data holds,

besides awake background EEG activity, the muscle

artifacts as well as 60 Hz power line interference.

3.1 Experimental Setup

Fig.1 summarizes the experimental outline proposed:

a) EEG is modeled by EDS and pre-processed follow-

ing (De Clercq et al., 2005); b) EEG rhythms are ob-

tained by Stochastic Relevance Analysis (SRA) and

RCE method; and c) Classiﬁer performance and rele-

vant channels are obtained to determine the presence

of epilepsy. The selected channels will be evaluated

both visual inspection by expert neurologist and elec-

trode placement on topographic map.

Raw EEG

Modelling and

pre-processing:

EDS

Feature extraction:

RCE and

Stochastic analysis

k-NN

classifier

Classification

Channel

selection

Figure 1: A general scheme with the proposed method for

EEG classiﬁcation and scalp localization of seizure activity.

EDS model uses four parameters: d determin-

ing which left singular vectors belong to the signal

subspace, N samples, and Hankel matrix dimension

L×M. The values used here corresponds to the values

suggested in (De Clercq et al., 2005) were N = 30720,

L = N/2 = 15371, M = L − 1 = 15370 and d = 0.01.

It was observed by visual inspection that this method

successfully had removed muscle artifact while the

epileptic activity had been enhanced.

In this work, we attempt to extract the frequency

band between 0.5 and 8 Hz, which is the most closely

related to epilepsy (Nam et al., 2002; Guerrero-

Mosquera et al., 2010). All experiments proposed

in this work use real and multi-channel EEG data

and have been supervised by an expert neurologist.

Regarding the relevance measure, both RCE and

stochastic variability are used to ﬁnd the EEG chan-

nels with greater relevance weight in presence of

epileptic seizure and then ﬁnding approximately the

damaged brain region.

Each rhythm is used as a dynamic feature for the

classiﬁer training. After obtaining the feature matrix,

Principal Component Analysis (PCA) is used as a fea-

ture extraction method to reduce the high dimension-

ality of the feature matrix. The number of principal

components (PCs) is selected based on the number of

PCs that maximizes the performance measures in the

classiﬁer. Cross-validation procedure is used to eval-

uate the performance of proposed experiments, which

consists in dividing the database into 10 folds, each

one with an equal number of signals per class. A

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

286

k− nearest neighbors (k-nn) classiﬁer is trained with

k = 5. The classiﬁcation performance is measured by

means of the accuracy, sensitivity and speciﬁcity, de-

ﬁned by:

(%) =

∗ 100

Sens(%) =

T P

+ N

∗ 100;

Spec(%) =

T N

+ N

∗ 100;

where N

is the number of correctly classiﬁed pat-

terns, N

is the total number of patterns used to feed

the classiﬁer, N

T P

is the number of true positives (ob-

jective class accurately classiﬁed), N

is the number

of false negatives (objective class classiﬁed as refer-

ence class), N

T N

is the number of true negatives (ref-

erence class classiﬁed as objective class), and N

the number of false positives (reference class classi-

ﬁed as objective class).

Table 1: Accuracy values for each EEG channel by Stochas-

tic Relevance Analysis (SRA) for all patients (DB1). Values

in bold type are those selected by neurologist through a vi-

sual inspection. Note the correspondence between high val-

ues obtained by SRA and EEG channels chosen by visual

inspection.

EEG

Patient

EEG

Patient

EEG

Patient

1 2 3 4 5 6

F4 0.16 0.06 Fp1 0.19 F4 0.59 0.27 0.66

FP2 0.46 0.06 F3 0.08 Fp2 0.14 0.13 0.49

F3 0.43 0.20 C3 0.11 F3 0.32 0.11 0.40

FP1 0.18 0.36 P3 0.05 Fp1 0.18 0.15 0.41

T6 0.31 0.09 O1 0.04 T6 0.50 0.24 0.31

T5 0.98 0.15 F7 0.06 T5 0.20 0.14 0.36

O2 0.48 0.45 T3 0.36 O2 0.11 0.01 0.29

O1 0.62 0.31 T5 0.19 O1 0.26 0.07 0.20

F7 0.82 0.32 Fp2 0.17 F7 0.77 0.11 0.39

F8 0.58 1.00 F4 0.11 F8 0.25 0.48 0.74

T3 0.98 0.15 C4 0.46 T3 0.13 0.70 0.32

T4 0.22 0.75 P4 0.18 T4 0.89 0.05 0.84

C4 0.51 0.47 O2 0.16 C4 0.20 0.03 0.20

C3 0.35 0.15 F8 0.32 C3 0.21 0.06 0.21

P4 0.75 0.16 T4 1.00 P4 0.27 0.23 0.25

P3 0.59 0.12 T6 0.88 P3 0.07 0.15 0.29

Cz 0.77 0.42 Fz 0.17 Cz 0.69 0.14 0.15

Ecgp 0.11 0.01 Cz 0.15 Ecgp 0.6 0.01 0.08

Pz 0.85 0.59 Pz 0.13 Pz 0.40 0.18 0.38

T1 1.00 0.15 A1 0.10 T1 1.00 1.00 0.13

T2 0.29 0.63 A2 0.72 T2 0.08 0.09 1.00

Fz 0.47 0.53 Ecg1 0.02 A1 0.84 0.91 0.59

Ecgn 0.04 0.03 T1 0.21 A2 0.34 0.12 0.82

T2 0.13 Fz 0.25 0.22 0.62

Ecgn 0.01 0.01 0.07

4 RESULTS

This section shows the effectiveness of the methodol-

ogy proposed by comparing the approach with Rhyth-

mic Component Extraction (RCE) in two sceneries:

channel selection and EEG classiﬁcation.

4.1 Channel Selection

Table 1 shows the estimated stochastic relevance val-

ues for each EEG channel (DB1 data), computed by

SRA. Values in bold type are those selected by neu-

rologist through a visual inspection. Note the corre-

spondence between high values obtained by SRA and

EEG channels chosen by visual inspection.

Localization of channels with high relevance values in single EEG data

(Pacient 1)

Channels by visual inspection

(T1, T3 and T5)

Channels by stochastic relevance

(T1, T3 and T5)

Channels by RCE

(FP1,F3 and P4)

F4F´2F3F´1T6 T5O2O1F7F8T3T4C4C3P4 P3CzEpPz T1T2 Fz En

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

EEG Channels

Normalized Amplitude w

SRA

RCE

Figure 2: Comparison of the channels selected by RCE,

stochastic relevance (SRA) and visual inspection at patient

1 (DB1 database). Upper: Localization of the channels se-

lected on scalp topographic map. Black areas represent a

highest energy concentration than grey areas. Bottom: Rel-

evance values obtained by RCE method and stochastic rel-

evance for each channel. Note that EEG channels with high

weighting by stochastic relevance (bottom) correspond to

channels located on epileptogenic region (upper).

Fig. 2 (upper) shows corresponding channels selected

by RCE and stochastic relevance (SRA) with the vi-

sual inspection on scalp topographic map at patient

1 (DB1). It can be seen that high values achieved by

RCE method do not correspond to channels chosen by

the neurologist through visual inspection. The Fig. 2

AChannelSelectionMethodforEEGClassificationbasedonExponentiallyDampedSinusoidalModelandStochastic

RelevanceAnalysis

287

(bottom) depicts relevance values obtained for each

DB1 channel at the patient 1 by RCE method and

stochastic relevance. Note that SRA measures suc-

cessfully selects the channels with more seizure ac-

tivity (T1, T3 and T5) showing high relevance values

close to 1. All results show that our method provide

a better medical interpretability about epileptic region

compared to the weights computed by RCE method.

Shortest weights values corresponding to EEG deriva-

tions are discarded in this ﬁgure.

4.2 EEG Segment Classiﬁcation

Table 2 shows classiﬁcation results achieved in DB1

and DB2 databases by RCE and stochastic variability

(SRA). As seen from Table 2, there is an improve-

ment in terms of classiﬁer performance when we use

stochastic relevance, DB1 achieves 95.38% and DB2

achieves 94.60% in accuracy values, compare to RCE

that present accuracy values for DB1 (93.60%) and

DB2 (92.10%).

Table 2: Comparison in classiﬁcation performance of RCE

and SRA by method by k-NN algorithm.

Method DB Acc( %) Sens(%) Spec(%)

RCE

DB1 93.60 93.60 93.60

DB2 92.10 93.85 90.35

SRA

DB1 95.38 95.90 96.23

DB2 94.60 94.60 94.32

Note that values in classiﬁer performance are

close each to other, even though DB2 is more con-

taminated by artifacts than DB1, which has a pre-

processing to eliminate ocular movements (Guerrero-

Mosquera and Navia-Vazquez, 2012). The proposed

model is then stable for EEG classiﬁcation problems

in presence of noise.

5 CONCLUSIONS AND FUTURE

WORK

A method for classiﬁcation and channel selection for

EEG multi-channel data with epilepsy is proposed.

This method, based on Exponentially Damped Sinu-

soidal model (EDS) and Stochastic Relevance Anal-

ysis (SRA), is simple and do not requires high com-

putational cost in training than others methods pro-

posed at the state-of-the-art (Peters et al., 2001; Yang

et al., 2012; Duun-Henriksen et al., 2012). Achieved

results show this method is an alternative for extract-

ing relevant EEG rhythms and selecting EEG chan-

nels with epileptic activity. Results validated by ex-

perts through visual inspection and scalp topographic

map, show that the approach also provides a better

medical support in epileptic region localization.

Future work includes: comparing our approach with

other seizure detection methods proposed in the state-

of-the-art; exploring other brain abnormalities such

as Alzheimer, sleep disorders and dementia; exploit-

ing the features for epileptogenic region analysis, and

consider the feasibility of our method to seizure an-

ticipation.

ACKNOWLEDGEMENTS

This research is carried out under the grant Centro de

Investigaci

on e Innovaci

on de Excelencia ARTICA,

sponsored by COLCIENCIAS and Convocatoria de

apoyo a tesis de posgrado DIMA 2011, Universidad

Nacional de Colombia, project 13753. Also thanks

to Instituto de Epilepsia y Parkinson del Eje Cafetero

(Pereira,Colombia) with the EEG data collection.

REFERENCES

De Clercq, W., Vanrumste, B., Papy, J.-M., Van Paesschen,

W., and Van Huffel, S. (2005). Modeling common

dynamics in multichannel signals with applications

to artifact and background removal in EEG record-

ings. IEEE Transactions on Biomedical Engineering,

52(12):2006–2015.

Duun-Henriksen, J., Kjaer, T. W., and et al (2012). Chan-

nel selection for automatic seizure detection. Clinical

Neurophysiology, 123:84–92.

Guerrero-Mosquera, C., Malanda-Trigueros, A., Iriarte-

Franco, J., and Navia-Vazquez, A. (2010). New fea-

ture extraction approach for epileptic EEG signal de-

tection using time-frequency distributions. Med. Biol.

Eng. Comput., 48:321–330.

Guerrero-Mosquera, C. and Navia-Vazquez, A. (2012). Au-

tomatic removal of ocular artifacts using Adaptive Fil-

tering and Independent Component Analysis for EEG

data. IET Signal Processing, 6:99–106.

Lodder, S. and Putten, M. V. (2011). Automated eeg anal-

ysis: Characterizing the posterior dominant rhythm.

Journal of Neuroscience Methods , 200:86–93.

Nam, H., Yim, T.-G., Han, S. K., Oh, J.-B., and Lee,

S. K. (2002). Independent Component Analysis of ic-

tal EEG in medial temporal lobe epilepsy. Epilepsia,

43:160–164.

Orekhova, E., Elam, M., and Orekhov, V. (2011). Adaptive

estimation of EEG-rhythms for optimal band identi-

ﬁcation in BCI. Journal of Neuroscience Methods,

195:47–60.

Peters, B., Pfurtscheller, G., and Flyvbjerg, H. (2001). Au-

tomatic differentiation of Multichannel EEG signals.

IEE Trans. On Bio Med Eng, 48:111–116.

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

288

Sepulveda-Cano, L., Acosta-Medina, C., and Castellanos-

Dominguez, G. (2011). Relevance Analysis of

Stochastic Biosignals for Identiﬁcation of Patholo-

gies. EURASIP Journal on Advances in Signal Pro-

cessing, 2011:10.

Sri, K. S. and Rajapakse, J. C. (2008). Extracting EEG

rhythms using ICA-R. IEEE International Joint Con-

ference on Neural Networks, pages 2133–2138.

Tanaka, T. and Saito, Y. (2008). Rhythmic component ex-

traction for multi-channel eeg data analysis. In Acous-

tics, Speech and Signal Processing, 2008. ICASSP

2008. IEEE International Conference on, pages 425–

428.

Veluvolu, K., Wang, Y., and Kavuri, S. (2012). Adaptive

estimation of EEG-rhythms for optimal band identi-

ﬁcation in BCI. Journal of Neuroscience Methods,

212:163–172.

Yang, J., Singh, H., and et al. (2012). Channel selection and

classiﬁcation of electroencephalogram signals: An ar-

tiﬁcial neural network and genetic algorithm-based

approach. Artiﬁcial Intelligence in Medicine, 55:117–

126.

Zalay, O., Kang, E., Cotic, M., Carlen, P., and Bardakjian,

B. (2009). A wavelet packet-based algorithm for the

extraction of neural rhythms. Annals of Biomed. Eng.,

37:595–613.

AChannelSelectionMethodforEEGClassificationbasedonExponentiallyDampedSinusoidalModelandStochastic

RelevanceAnalysis

289