Classification of Normal and Epileptic Seizure EEG Signals Based on Empirical Mode Decomposition - Complex System Modelling and Control Through Intelligent Soft Computations - page 377

Table 1 Classi cation
performance of the proposed
method for different kernel
functions

Performance parameters Linear Polynomial

(d ¼ 2)

RBF
( r ¼ 1)

SEN (%) 100 100 100

SPF (%) 97.00 90.00 100

ACC (%) 98.50 99.00 100

PPV (%) 97.69 90.91 100

NPV (%) 100 100 100

MCC 0.97 0.98 1

Along with area parameter of analytic IMFs, we have also computed 95 % confi- -

dence ellipse area of SODP for first four IMFs of EEG signals which covers around

95 % points in SODP. By considering both area measures for first four IMFs, lead

to eight features that forms the final input feature set for LS-SVM based classifi- -

cation of normal and epileptic seizure EEG signals.
SVM is a supervise machine learning approach, suitable for small-sample dataset
(Azar and El-Said 2014 ). LS-SVM is the least square reformulation of the SVM
problem (Suykens and Vandewalle 1999 ) which uses equality constraints, instead of
inequality constraint used in standard SVM. Consequently, solution follows from set
of linear equations instead of quadratic programming problem. Hence, LS-SVMoffers
less computational complexity with excellent generalised performance (Suykens and
Vandewalle 1999 ). In thiswork, the area parameters computed from the IMFs has been
used as input feature set for LS-SVM classi er for classi cation of normal and epi-

leptic seizure EEG signals. In order to evaluate the classi cation performance, dif-

ferent kernel functions have been utilized and their performance parameter values
have been shown in Table 1 . Various performance parameters discussed in previous
section have been computed for three kernel functions which are linear kernel,
polynomial kernel, and radial basis function (RBF) kernel. It can be observed that
performance parameter values for RBF kernel are best among three kernel functions.
The value of scaling factor associated with RBF kernel has been set empirically as 1.
The ten-fold cross validation procedure is suitable for evaluating classi

cation

accuracy of a classi cation of biomedical signals (Sharma et al. 2014 ;

Pachori and Patidar 2014 ). In this study, ten-fold cross validation procedure has been
employed to evaluate the classi

er for classi

cation performance of LS-SVM classi er.

cation accuracy achieved using proposed method with RBF kernel is
100 % which suggests successful identi

The classi

cation of all, normal and epileptic seizure
EEG signals. The resulting 100 % sensitivity shows the correct identi cation of all

epileptic seizure EEG signals and 100 % speci cation

by not recognizing any normal EEG signal as epileptic seizure EEG signal.
Moreover, Table 2 shows the results obtained with proposed method and some
other existing methods using the same dataset. Different parameters analysed for
classi

city shows adequate classi

cation in other compared methods have also been mentioned in Table 2 .It
should be noted that the performance of the proposed method in terms of classi-
fication accuracy is same as that of discussed in Tzallas et al. ( 2007 ), in which time-
frequency analysis based parameters have been used for classi cation. The area