Available online at www.sciencedirect.

com

Expert Systems
with Applications
Expert Systems with Applications 36 (2009) 13291336
www.elsevier.com/locate/eswa

Cross-correlation aided support vector machine classier

for classication of EEG signals
Suryannarayana Chandaka, Amitava Chatterjee *, Sugata Munshi
Department of Electrical Engineering, Jadavpur University, Kolkata 700 032, India

Abstract
Over the last few decades pattern classication has been one of the most challenging area of research. In the present-age pattern classication problems, the support vector machines (SVMs) have been extensively adopted as machine learning tools. SVM achieves higher
generalization performance, as it utilizes an induction principle called structural risk minimization (SRM) principle. The SRM principle
seeks to minimize the upper bound of the generalization error consisting of the sum of the training error and a condence interval. SVMs
are basically designed for binary classication problems and employs supervised learning to nd the optimal separating hyperplane
between the two classes of data. The main objective of this paper is to introduce a most promising pattern recognition technique called
cross-correlation aided SVM based classier. The idea of using cross-correlation for feature extraction is relatively new in the domain of
pattern recognition. In this paper, the proposed technique has been utilized for binary classication of EEG signals. The binary classiers
employ suitable features extracted from crosscorrelograms of EEG signals. These cross-correlation aided SVM classiers have been
employed for some benchmark EEG signals and the proposed method could achieve classication accuracy as high as 95.96% compared
to a recently proposed method where the reported accuracy was 94.5%.
2007 Elsevier Ltd. All rights reserved.
Keywords: Support vector machines; Cross-correlation; Electroencephalogram signals; Optimal separating hyperplane

1. Introduction
Epileptic seizures are due to temporary electrical disturbances of the brain. Unfortunately, the occurrence of epileptic seizure is unpredictable (Subasi, 2007). The
neuroscientists and biological psychiatrists prefer to study
the electrical activity of the brain with the help of electroencephalographic records for diagnosis of neurological disorders (Gular, Ubeyli, & Guler, 2005). Basically, the
electroencephalogram (EEG) is a complex and aperiodic
time series containing information of the electrical activity
generated by cerebral cortex nerve cells. Earlier the analysis
of the electroencephalograph (EEG) records was restricted
to visual inspection. It has been shown that such visual
inspection is insucient to closely study the minute obser-

Corresponding author.
E-mail address: cha_ami@yahoo.co.in (A. Chatterjee).

0957-4174/$ - see front matter 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2007.11.017

vational variations in EEG signals, thus necessitating

computer based automation tools (Subasia, Alkana, Koklukayab, & Kiymika, 2005).
Spectral analysis can be recommended for EEG signals,
which can give information regarding brain activities.
Nowadays articial neural networks (ANNs) may oer a
superior performance for analysis of EEG signals, compared to the spectral analysis methods (Guler, Ubeyli, &
Guler, 2005). As compared to the conventional spectral
analysis methods, ANNs can eectively make a decision
regarding the class of the signal. Neural networks have
been successfully adopted for so many medical applications
(Baxt, 1990). The invention of neural network models produces a great revolution in the area of pattern recognition.
Support vector machines constitute one of the relatively
contemporary branches of neural network models. Nowadays in the machine learning technologies for pattern classication problems, the support vector machines have been
extensively adopted for machine learning tools. In contrast

1330

S. Chandaka et al. / Expert Systems with Applications 36 (2009) 13291336

to traditional neural networks SVM achieves higher generalization performance, due to the utilization of structural
risk minimization principle. The SRM principle seeks to
minimize the upper bound of the generalization error consisting of the sum of the training error and a condence
interval, where as incase of traditional neural networks
the empirical risk minimization principle only minimizes
the training error (Cao & Tay, 2003).
Several works on the classication of EEG signals have
been reported. In Guler et al. (2005), the researchers evaluated the classication accuracy of the recurrent neural
networks (RNNs) using Lyapunov exponents trained with
LevenbergMarquardt algorithm on the electroencephalogram (EEG) signals. The database used for their work is
EEG database (EEG time series), which is developed by
department of Epileptology, University of Bonn, California. In this work three classes of EEG signals have been
considered. They are from healthy volunteers with eyes
open, epilepsy patients in the epileptogenic zone during a
seizure-free interval, and epilepsy patients during epileptic
seizures. An overall classication accuracy of 96.79% was
obtained in this work. In Subasi (2007), mixture of experts
model (ME) was utilized for classication of EEG signals.
A double-loop expectation-maximization (EM) algorithm
was introduced to the ME network structure for detection
of epileptic seizure and wavelet feature extraction was
employed. The database used in this work was with University of Bonn EEG time series, California. Healthy volunteers with eyes open and epilepsy patients during
epileptic seizures were used for classication and overall
accuracy was 94.5%. Up to now no study has been reported
in literature related to cross-correlation based feature
extraction in the analysis of EEG signals.
Correlation is a mathematical operation that is very similar to convolution. Cross-correlation of two signals measures the extent of similarity between these signals. In the
present work, a healthy subject is considered as reference
and the EEG signal of the reference subject is correlated
with the EEG signal of each of the other subjects. From
each of resulting cross-correlation sequences a set of ve
features is extracted. These feature vectors called patterns
are segregated into training and testing dataset and the
training set is utilized for training SVM classier. The
LS-SVM package, available in svm toolbox has been used
for this purpose. The generalization capability of the classier is then tested by utilizing it for testing dataset. The
structure of the paper is organized as follows. Section 2
briey reviews the sets of the EEG signals used in our
work. Section 3 reviews the cross-correlation of discretetime signals and the computation of various features of
crosscorrelograms utilized in this work. Section 4 is
devoted to provide necessary background and basic idea
behind support vector machine. Section 5 presents how
the proposed method is applied for classication of EEG
signals and how the cross-correlation is performed and features are extracted for each EEG signals is explained in
detail. Finally how the SVM is utilized for classication

is discussed. Chapter 6 highlights the conclusions of the

work and discuses directions for future investigations.
2. EEG signals
In this work, a publicly available database (EEG time
series; Andrzejak et al., 2001) has been used. All EEG signals recorded were with the same 128-channel amplier system, using an average common reference. The data were
digitized at 173.61 samples per second using 12-bit resolution. Band-pass lter settings were 0.5340 Hz (12 dB/
oct). The EEG segments in this database were cut out from
continuous multi-channel EEG recordings after visual
inspection for artifacts, e.g., due to muscle activity or eye
movements. Fig. 1 depicts the electrode placement for
recording of EEG signal. The complete dataset consists of
ve sets (denoted AE), each containing 100 single-channel
EEG signals of 23.6 s. Fig. 2 depicts example of EEG signals of each of the ve sets. Sets A and B have been taken
from surface EEG recordings of ve healthy volunteers with
eyes open and closed, respectively. Signals in two sets have
been measured in seizure-free intervals from ve patients in
the epileptogenic zone (D) and from the hippocampal formation of the opposite hemisphere of the brain (C). Set E
contains seizure activity, selected from all recording sites
exhibiting ictal activity. Sets A and B have been recorded
extra cranially, whereas sets C, D, and E have been recorded
intracranially. In the present study we classied only two (A
and E) of the complete dataset.
3. Cross-correlation of signals
Correlation is a mathematical operation that is very similar to convolution. In correlation, a cross-correlation
sequence between two energy signals measures the extent
of similarity between these two signals (Proakis & Manolakis, 1997). If a signal is correlated with itself, the resulting

Fig. 1. The 1020 electrode placement for recording a EEG pattern.

S. Chandaka et al. / Expert Systems with Applications 36 (2009) 13291336

1331

Set A

100
0

Set B

-100
0
500

500

1000

1500

2000

2500

3000

3500

4000

4500

500

1000

1500

2000

2500

3000

3500

4000

4500

500

1000

1500

2000

2500

3000

3500

4000

4500

500

1000

1500

2000

2500

3000

3500

4000

4500

500

1000

1500

2000

2500

3000

3500

4000

4500

Set C

-500
-5
0
200
0

Set D

-200
0
500
0
-500
Set E

1000
0
-1
-1000

Fig. 2. Exemplary EEG time series from each of the ve sets. Amplitude of surface EEG recordings are typically in the order of micro volts.

sequence is called the auto-correlation sequence. Suppose

that we have two signal sequences x(n) and y(n) each of
which has nite energy. The cross-correlation of x(n) and
y(n) is given by
8
N m1
>
< P xnm y m P 0
n
b xy m
1
R
n0
>
:b
m<0
R yx m
where m = . . .2, 1, 0, 1, 2,. . .. The index m represents
time shift parameter also known as lag and subscript xy
represents sequences being correlated. The order of the
subscripts, with x preceding y, indicates the direction in
which one sequence is shifted, relative to other. If each of
the signals x(n) and y(n) has nite number of samples M,
then the resultant cross-correlation sequence has 2M  1
no of samples. Let us assume that Fig. 3 depicts a typical
crosscorrelogram of a signal. In this proposed approach
ve features are extracted from each crosscorrelogram.
The features are the peak value, instant at which peak ocR

xy

(m)

-M

Fig. 3. Example of a typical crosscorrelogram.

curs, centroid, equivalent width and mean square abscissa.

Let the crosscorrelated sequence be represented by R(m).
Peak value and instant at which peak occurs can be easily
determined from crosscorrelogram. The rest of the three
features centroid, equivalent width and mean square abscissa can be computed as follows
PM
mRm
2
Centroid PmM
M
mM Rm
PM
mM Rm
Equivalent width
3
Peak value of Rm
PM
2
mM m Rm
4
Mean square abscissa P
M
mM Rm
4. Support vector machines
Support vector machine is a contemporary machine
learning tool, which has quickly evolved as an active area
of research because of its strong theoretical foundation
(Kumar, 2004). SVMs were invented from Vapniks statistical learning theory (Vapnik, 1995). The SVM is specically formulated to solve a binary classication problem
in a supervised manner and the learning problem is formulated as a quadratic optimization problem where the error
surface is free of any local minimum and has global optimum (Begg, Palaniswami, & Owen, 2005). In case of binary
classication problems the main objective of SVM is to
build an optimal separating hyper plane (OSH) in such a
way that the margin of separation between two classes is
maximized (Haykin, 1999). The machine achieves this
desirable property on the basis of the principle of structural
risk minimization principle. The SRM principle seeks to

1332

S. Chandaka et al. / Expert Systems with Applications 36 (2009) 13291336

minimize the upper bound of the generalization error consisting of the sum of the training error and a condence
interval (Cao & Tay, 2003).
To develop the SVM based classiers for linearly separable patterns, let us consider a training set represented
N
by fxi ; d i gi1 where xi is the n-dimensional input feature
vector and di represents the target output. The input patterns represented by the target output di = 1 constitute
the positive group and the target output di = 1 constitute
the negative group. For linearly separable case it is necessary to nd a hyperplane to distinguish the two classes of
input vectors (Haykin, 1999; Foody & Mathur, 2004).
There may be so many hyperplanes to separate the input
patterns, but there will be one and only one optimal separating hyperplane (OSH). The optimal separating hyperplane is a hyperplane that passes between two classes in
such a way that the distance between the closest training
points and itself is as maximum as possible.
The equation of decision surface in the form of hyperplane is written as w  x + b = 0, where x is the input feature vector, w is the adjustable weight vector and b is the
bias. For linearly separable case, a separating hyperplane
can be found for two classes such that w  x + b P +1 for
positive group of data and w  x + b 6 1 for negative
group of data. These two equations can be combined
(Osuna, Freund, & Girosi, 1997; Kumar, 2004) as
d i w  x b  1 P 0

The distance from origin to the optimal hyper plane is

(kbk)/(kwk) and kwk is the Euclidean norm of w (Burges,
1998). Fig. 4 shows that for some data points (xi, di),
w.x + b = + 1 is satised and for some other data points
(xi, di), w.x + b = 1 is satised. These points constitute
support vectors and hence the method is called support
vector machine. The support vectors denote those data
points, which are closest to the OSH, and they are most difcult to classify (Haykin, 1999). The goal is to maximize

the margin of separation (q) and hence to determine wo

and bo for the OSH. The distance between two hyper
planes is 2/(kwk) (Burges, 1998; Gunn, 1998). Hence the
constrained optimization problem can be stated as (HayN
kin, 1999): For the given training samples fxi ; d i gi1 , nd
optimum w and b subjected to the constraints of Eq. (5)
(Foody & Mathur, 2004; Osuna et al., 1997; Gunn, 1998).
minimize
subject to

1
2
uw kwk
2
d i w  x b P 1;

The above constrained optimization problem is solved by

using Lagrangian multiplier method. In practice, we mostly
encounter patterns, which are linearly non-separable, i.e. the
data points of dierent classes do overlap. The constraints
dened in Eq. (5) cannot be satised for all the data points.
So, in order to deal with such situations, it is necessary to
N
introduce a new set of slack variables fni gi1 , that measures
the amount of violation of constraints (Osuna et al., 1997).
Then the optimization problem can be reformulated as
(Chang & Lin, 2001; Osuna et al., 1997) (see Fig. 5)
!k
N
X
1
2
ni
minimize uw; N kwk C
2
i1
7
subject to d i w  x b P 1  ni ; i 1; 2; . . . ; N
ni P 0; i 1; 2; . . . N
where C is called the regularization parameter. The second
part of the objective function of the Eq. (7) seeks to penalize the data points located in the incorrect side of the decision boundary. If the data points overlap considerably in
feature space, then the penalty term becomes very large,
and the hyperplane may not generalize well (Foody &
Mathur, 2004).
The basic approaches discussed above can be extended
to nonlinear decision surfaces. In nonlinear decision sur1- 1

Class 1

Class 1

Support
vector

Origin

i 1; 2 . . . N

Support
vector

Origin
x1
x2

Margin of
Separation
Class 2

Support
vector

Fig. 4. The SVM for the linearly separable case. The middle line indicates
optimal separating hyper plane.

Margin of
Separation
Class 2

Support
vector

-1-- 2

Fig. 5. The SVM for the linearly non-separable case. The middle line
indicates optimal separating hyper plane.

i1

where ai, i = 1, 2, . . . N are called Lagrange multipliers. For

train the SVM classier only the kernel is required. Gaussian radial basis kernel is one of the such kernel and is given
by
Kx; y e

kxyk2
 2
2r

150
100
50

--100
-150
-200
0

500 1000 1500 2000 2500 3000 3500 4000 4500

samples

Fig. 6a. Reference EEG segment.

200
150
100
50
0
-50
-100
-150
-200

where r2 is kernel parameter or width. The values chosen for

the two kernel parameters (C, r2) signicantly aect the classication accuracy of SVM classier. Large values of (C, r2)
may lead to over tting problem for the training data
(Chang & Lin, 2001; Foody & Mathur, 2004). So the values
must be chosen carefully. Various kernels that are used frequently are given in Table 1. The type of kernel chosen for
training SVM classier must satisfy Mercers condition.

-250

500 1000 1500 2000 2500 3000 3500 4000 4500

samples

Fig. 6b. EEG segment of a healthy volunteer.

Amplitude 2 (V)2

The present method is employed to classify signals

belonging to Set A and Set E in EEG time series. In each
of Set A and E, there are 100 EEG segments i.e. 100 signals
each consisting of 4096 samples. Figs. 6b and 7b show typical EEG segments of healthy volunteer and epileptic
patient. We performed two dierent studies with dierent
sizes of training set and testing set. In the rst case study,
50% of original data for each class (100 time series of
4096 samples) were used for training (50 signals from each
class, each of 4096 samples) and remaining 50% for testing

200

-50

5. Investigations and results

1333

Amplitude (V)

faces, the same data points are transformed from input

space to the high dimensional feature space with some nonlinear mapping (Haykin, 1999; Osuna et al., 1997; Begg
et al., 2005). In this high dimensional feature space the
spreading of data points yields a linear hyperplane (Foody
& Mathur, 2004). Now suppose the training data is
mapped to some other (possibly innite dimensional)
Euclidean space H through a nonlinear function / such
that U:Rn ? H. So in this high dimensional feature space
the input vector x can be characterized as /(x) (Burges,
1998; Foody & Mathur, 2004). Then of course the training
algorithm would only depend on the data through dot
products in H i.e. on functions of the form /(x)  /(xi).
Now if there were a kernel function K such that
K(x, xi) = /(x)  /(xi), we would only need to use K in the
training algorithm and would never need to explicitly even
know what / is Burges (1998). Then the decision function
is given by Burges (1998), Osuna et al. (1997), Gunn (1998)
!
N
X
ai d i kx; xi b
8
f x sign

Amplitude (V)

S. Chandaka et al. / Expert Systems with Applications 36 (2009) 13291336

x10
1.5

1
0.5
0
-0.5
-1

-5000 -4000 -3000-2000-1000 0 1000 2000 3000 4000 5000

samples
Fig. 6c. Crosscorrelogram of the healthy volunteer.

Table 1
Various types of kernels
Type of classier

Kernel function

Polynomial of degree m
Multi layer perceptron
Radial basis function

K(x, y) = (1 + x  y)m
K(x, y) = tan h(x  y  h)
K(x, y) = exp(kx  yk2)

(50 signals from each class, each of 4096 samples) the classier. In the second case study, a rectangular-window was
formed by 512 samples, such that each EEG segment is further divided into eight small segments. Hence in each set
there are 800 segments.

1334

S. Chandaka et al. / Expert Systems with Applications 36 (2009) 13291336

Amplitude (V)

200
150
100
50
0
-50
-100
-150
-200

500 1000 1500 2000 2500 3000 3500 4000 4500

samples

Amplitude (V)

Fig. 7a. Reference EEG segment.

600
400
200
0

represent the reference signal. This reference data is crosscorrelated with the data of the rest of healthy volunteers
and all the epilepsy patients. So in this process we create
99 feature vectors for healthy set, and 100 feature vectors
for epileptic seizure set. Figs. 6c and 7c shows typical
results of cross-correlation for of healthy volunteer and
epilepsy patient. In the second case study there are eight
reference data sub segments and each such sub segment is
crosscorrelated with the corresponding segment of the
data, belonging to either healthy dataset or epileptic dataset. So there are 792 feature vectors for healthy set and
there are 800 feature vectors formed for the epileptic seizure set. For both the two experiments the extracted ve
features were peak value (PV), instant at which peak occurs
(IP), equivalent width (EW), centroid (C) and mean square
abscissa (MSA). Hence for each crosscorrelogram we have
created ve-dimensional feature vector [PV IP EW C
MSA]T. These feature vectors are segregated into training
and testing dataset as described before, and the training
set is utilized for training SVM classier. We utilized the
LS-SVM package, available in svm toolbox for this purpose. The generalization capability of the classier is then
tested by utilizing it for testing dataset.
Tables 2 and 3 describe the segregation of signal segments into training and testing datasets for the two studies.

-200
Table 2
Segregation of training and testing sets for case study 1

-400
-600

500 1000 1500 2000 2500 3000 3500 4000 4500

samples

Class

Training set

Testing set

Total

Healthy
Epileptic
Total

50
50
100

49
50
99

99
100
199

Fig. 7b. EEG segment of a seizure volunteer.

Amplitude 2 ((V)
V) 2

x 10

2
1

Table 3
Segregation of training and testing sets for case study 2
Class

Training set

Testing set

Total

Healthy
Epileptic
Total

492
500
992

300
300
600

792
800
1592

0
-1

Table 4
Confusion matrix for experiment 1
Predicted output

-2
-3

Actual output
-5000 -4000-3000-2000-1000 0 1000 2000 3000 4000 5000
samples

Fig. 7c. Crosscorrelogram of a seizure patient.

To perform the classication, at rst cross-correlation is

employed for feature extraction (Proakis & Manolakis,
1997). In the rst case study undertaken data for one
healthy person is considered as reference. Figs. 6a and 7a

Epileptic
Healthy

Epileptic

Healthy

46
0

4
49

Table 5
Confusion Matrix for experiment 2
Predicted output

Actual output

Epileptic
Healthy

Epileptic

Healthy

277
4

23
296

S. Chandaka et al. / Expert Systems with Applications 36 (2009) 13291336

Table 6
Comparison of statistical parameters of various classication algorithms
Type of
classier

ME
(Subasi,
2007)
model (%)

MLPNN
(Subasi,
2007) (%)

Crosscorrelation
SVM
(Experiment1)
(%)

Crosscorrelation
SVM
(Experiment2)
(%)

Sensitivity
Specicity
Accuracy

95.0
94.0
94.5

93.6
92.6
93.2

92.00
100.00
95.96

92.4
98.6
95.5

There are two important parameters, which should be

properly tuned for achieving desired performance. These
are optimal regularization parameter (C) and optimal kernel parameter (r2). For choosing these optimum values
cross-validation (Chang & Lin, 2001; Foody & Mathur,
2004) approach is highly recommended. In LS-SVM package L-fold cross-validation is provided for tuning hyper
parameters. Then the optimum values of C and r2 are
54.59 and 17.70, respectively. With these parameters chosen, the SVM is trained and the confusion matrices given
in Tables 4 and 5, for case studies 1 and 2, respectively
show the test performance. It can be seen that case study
1 produces an average accuracy of 95.96% and case study
2 produces an accuracy of 95.50%. As per the confusion
matrix of experiment 2, SVM misclassied four healthy
segments as epileptic seizure segments and 23 epileptic seizure segments as healthy segments. Similarly as per the
confusion matrix of case study 1 SVM classied all the
healthy segments as healthy segments and 4 epileptic segments are misclassied as healthy segments. The test performance of the SVM was also determined by computing
the following statistical parameters (Guler et al., 2005;
Ubeyli & Guler, 2004):
Sensitivity: number of true positive decisions/number of
actually positive cases.
Specicity: number of true negative decisions/number of
actually negative cases.
Total classication accuracy: number of correct decisions/total number of cases.
Table 6 gives comparative study of the quantitative performance achieved by employing our proposed algorithm,
vis-a-vis other algorithm very recently reported in Subasi
(2007), employing wavelet analysis, mixture of expert
model and multi-layer perceptron neural network. It can
be seen that our proposed method could produce better
classication accuracy compared to those methods. Among
the two case studies, case study 1 produces slightly better
performance than case study 2.
6. Conclusions and future scope
This work presents an attempt to develop a general-purpose signal classication scheme, which can be utilized for
classifying dierent categories of one-dimensional signal.

1335

The classication scheme utilized crosscorrelogram based

feature extraction procedure from signals, which can be utilized to develop SVM based classiers. Diagnosing epilepsy
is not an easy task, which requires acquisition of the EEG of
the patient and gathering additional clinical information. A
cross-correlation aided SVM that classies subjects as having or not having an epileptic seizure provides a valuable
diagnostic decision support tool for physicians treating
potential epilepsy, since diering types of epileptic seizures
require dierent treatments. The proposed system utilizing
the SVM classier could achieve accuracy as high as
95.96% compared to a recently proposed method where
the reported accuracy was 94.5%. Hence the support vector
machine has the inherent ability to solve a pattern recognition tasks in a manner close to the optimum for the problem
of interest. Our approach adopted, not only gave better classication accuracy but also quite simple from implementation point of view. A constraint on SVM has been its
binary nature. In order to extend the use of SVM to the multi
class classication problems it is necessary to construct multiple binary classiers. Then a tertiary classier is required to
combine the binary classiers. We also intended to apply this
new scheme to multi class classication problems. For this
purpose we will consider healthy volunteers with eyes open,
epilepsy patients in the epileptogenic zone during a seizurefree interval, and epilepsy patients during epileptic seizures.
References
Andrzejak, R. G., Lehnertz, K., Mormann, F., Rieke, C., David, P., &
Elger, C. E. (2001). Indications of nonlinear deterministic and nite
dimensional structures in time series of brain electrical activity:
Dependence on recording region and brain state. Physical Review E,
64, 061907.
Baxt, W. G. (1990). Use of articial neural network for data analysis in
clinical decision making: The diagnosis of acute coronary occlusion.
Neural Computation, 2, 480489.
Begg, R. K., Palaniswami, M., & Owen, B. (2005). support vector
machines for automated gait classication. IEEE Transactions on
Biomedical Engineering, 52(5), 828838.
Burges, C. J. C. (1998). A tutorial on support vector machines for pattern
recognition. Data Mining Knowledge Discovery, 2(2), 121167.
Cao, L. J., & Tay, F. E. H. (2003). Support vector machine with adaptive
parameters in nancial time series forecasting. IEEE Transactions on
Neural Networks, 14(6).
Chang, C. -C. & Lin, C. -J. (2001). LIBSVM: a library for Support Vector
Machines. <http://www.csie.ntu.edu.tw/~cjlin/libsvm>.
EEG time series are available under <http://www.epileptologie-bonn.de/
front_content.php?idcat=193&lang=3&changelang=3>.
Guler, N. F., Ubeyli, E. D., & Guler, I. (2005). Recurrent neural networks
employing Lyapunov exponents for EEG signals classication. Expert
Systems with Applications, 29, 506514.
Foody, G. M., & Mathur, A. (2004). A relative evaluation of multiclass
image classication by support vector machines. IEEE Transactions on
Geoscience and Remote Sensing, 42(6), 13351343.
Gular, I., Ubeyli, E. D., & Guler, N. F. (2005). A mixture of experts
network structure for EEG signals classication. In IEEE engineering
in medicine and biology 27th annual conference, Shanghai, China.
Gunn, S. R. (1998). Support vector machines for classication and
regression. Image speech and intelligent system group, Department of
Electrical Computer Science, Southhampton, UK University of
Southampton, Technical report.

1336

S. Chandaka et al. / Expert Systems with Applications 36 (2009) 13291336

Haykin, S. (1999). Neural networks: A comprehensive foundation (2nd ed.).

Newyork: Pearson Education.
Kumar, S. (2004). Neural networks: A class room approach. New Delhi,
India: Tata McGraw-Hill Publishing Company Limited.
Osuna, E., Freund, R., & Girosi, F. (1997). Training support vector
machines: An application to face detection. In Proceedings of computer
vision and pattern recognition (pp. 130136).
Proakis, J. G., & Manolakis, D. G. (1997). Digital signal processing:
Principles algorithms and applications (3rd ed.). Newdelhi, India:
Prentice-Hall.
Subasi, A. (2007). EEG signal classication using wavelet feature
extraction and a mixture of expert model. Expert Systems with
Applications, 32, 10841093.

Subasia, A., Alkana, A., Koklukayab, E., & Kiymika, M. K. (2005).

Wavelet neural network classication of EEG signals by using AR
model with MLE preprocessing. Neural Networks, 18, 985997.
svm toolbox <http://www.esat.kuleuven.ac.be/sista/lssvmlab/>.
Ubeyli, E. D., & Guler, I. (2004). Detection of electrocardiographic
changes in partial epileptic patients using Lyapunov exponents with
multilayer perceptron neural networks. Engineering Applications of
Articial Intelligence, 17, 567576.
Vapnik, V. (1995). The nature of statistical learning theory. New York:
Springer-Verlag.