Wrapper Subset Feature Selection for Optimal Feature Selection in Epileptic Seizure Signal Classification

Abstract

Epilepsy is diagnosed by assessing the brain signal using an electroencephalograph (EEG). The assessment relies on manual visual inspection, which required experience and years of training. A computer-aided diagnose system can help neurologists assess the EEG signal. This study explores the epileptic condition by decomposing EEG signals using three levels of wavelet packet decomposition (WPD). Three orders of Daubechies mother wavelets are used. Since EEG is a non-stationary biological signal, an entropy measurement using the Shannon entropy is used to extract the signals’ information. The next process is combining the features from all levels of the decomposed signals producing 14 number of features. This study reduces the number of features using the wrapper feature subset selection (WFSS) method. The searching algorithm used is the sequential backward (SBS) and forward (SFS) selection method. The multilayer perceptron neural network (MLPNN) is used for the classification method. The system achieves the highest accuracy of 91% by using seven number of features obtained from WPD(db2) + WFSS(SBS) + MLPNN. The minimum number of features is obtained using WPD(db16) + WFSS(SFS) + MLPNN, which produces six features. While the use of WFSS(SFS) in db16 produces six features with the highest increase of accuracy by 22%. This indicates that the use of WFSS can obtain an optimal number of features set and can improve the system’s performance.