Epileptic Seizure Detection from EEG Signals Using Best Feature Subsets Based on Estimation of Mutual Information for Support Vector Machines and Naïve Bayes Classifiers

The detection of epileptic abnormality from electroencephalogram (EEG) signals is achieved using pattern recognition system. In this work, to distinguish the epileptic seizure from normal EEG signal, pattern recognition is applied. Novel pattern recognition is studied with the conventional features which are extracted from discrete wavelet transform (DWT) sub-bands D3-D5 and A5 to detect epileptic seizure with support vector machine (SVM) and naive Bayes (NB) classifier for 14 varying combinations of set A to D with set E. The open source EEG data which is provided by University of Bonn, Germany, and Christian Medical College and Hospital (CMCH), India, are used in this work. Further, the DWT coefficients obtained from D3-D5 to A5 sub-bands increases the computational burden of the classifier. So, feature selection based on the estimation of mutual information theory is applied to the DWT coefficients to obtain significant features required for the classifier to get better accuracy with lesser computational burden. The results show that SVM performs well for 9 data sets and NB performs better for 4 varying data sets. However, SVM and NB performed equally well in differentiating normal open EEG and epileptic data and afford an accuracy of 100%. In using CMCH data, SVM provides 100% accuracy with only top 2 ranked features.