Optimal input features selection of wavelet-based EEG signals using GA

Abstract

We present a method of selecting optimal input features from wavelet coefficients of electroencephalogram (EEG) signals. A combination of genetic algorithm (GA) and artificial neural network (ANN) are used to select the relevant features. In this investigation, classification accuracy and the fraction of a number of features rejected per total features is used as the fitness function to be optimized. The mental tasks of EEG signals from six channels are decomposed into five levels using discrete wavelet transform (DWT) produces 24 sub-bands with 96 input features. The features used to describe each sub-band are average energy, standard deviation, kurtosis and skewness of the distribution. This optimal input features are classified into five classes of mental tasks. Two types of selection algorithms are compared i.e. roulette wheel selection (RWS) and stochastic universal sampling (SUS). Results show that 11 to 12 input features with average classification accuracy rate of 81% to 82% with RWS is achieved compared to 16 input features of the same accuracy when SUS is adopted. It can be concluded that RWS performs better than SUS in this study.