Classification of Normal and Epileptic Seizure EEG Signals Based on Empirical Mode Decomposition - Complex System Modelling and Control Through Intelligent Soft Computations - page 375

the actual negatives correctly. For example, percentage of normal EEG signals
correctly identi cation would result in

100 % sensitivity by detecting all epileptic seizure EEG signals correctly. It also
exhibits 100 % speci

ed as not having seizures. A perfect classi

city by not recognizing any normal EEG signal as epileptic
seizure signal. Positive predictive value is, the fraction of total positive patterns,
which represents the actual positive patterns (Azar and El-Said 2014 ). Accuracy of
classi cation is proportion of number of patterns which are correctly classi ed.

Similarly, negative predictive value is, the fraction of total identi ed negative

patterns, which represent actual negative patterns. Considering, TP and TN repre-
sent the total number of correctly identi ed true positive patterns and true negative

patterns respectively, along with FP and FN represents total number of false
positive patterns and false negative patterns, respectively. The sensitivity (SEN),
speci

city (SPF), accuracy (ACC), positive prediction value (PPV), negative pre-
diction value (NPV) of classi er can be de ned as (Azar and El-Said 2014 ):

TP

SEN ¼

FN

100 ð%Þ ð 29 Þ

TP þ

TN
TN

SPF ¼

FP

100 ð%Þ ð 30 Þ

þ

TP þ TN

ACC ¼

FN

100 ð%Þ ð 31 Þ

TP þ TN þ FP þ

TP

PPV ¼

FP

100 ð%Þ ð 32 Þ

TP þ

TN

NPV ¼

FN

100 ð%Þ ð 33 Þ

TN þ

Matthews correlation coef cient (MCC) is another parameter to measure clas-

si cation accuracy of

imbalanced positive and negative patterns in dataset (Azar and El-Said 2014 ).
Higher the value of MCC parameter, the better the classi

cation performance, which is the indication of classi

er performance (Yuan

et al. 2007 ). The MCC parameter can be de ned as follows (Yuan et al. 2007 ):

FP
ð TP þ FN Þð TP þ FP Þð TN þ FN Þð TN þ FP Þ

TP TN FN

MCC ¼ p ð 34 Þ

:

3 Experimental Results and Discussion

Main steps of proposed method include applying EMD on EEG signals to obtain
IMFs, computation of both area measures for

first four IMFs, extraction and for-
mation of feature set, training and testing of LS-SVM classi er. The proposed