Feature Selection in Brain Computer Interface Using Genetics Method

Present-day Brain Computer Interfaces (BCIs) determine user´s intent from electrophysiological signals. BCI systems could ensure an important new communication or control option for those with motor disabilities and give those without disabilities supplementary control channel or control channel for special circumstances. Motor imagery-based BCIs are realized with a large electrodes set and highly sophisticated spatial filtering methods or with reduced subject specific bipolar channels. Electrocardiogram (ECoG) recordings for neuroprosthetics ensure a mesoscopic abstraction level of brain function between micro wire single neuron recordings and electroencephalogram (EEG). ECoG BCI, success relies on ability to extract features from neural activity related to goal directed behavior. This study presents a BCI system and also a comparative study of feature selection methods to improve the classification efficacy. ECoG signals are pre-processed and features extracted using Wavelet Packet Tree. The features are selected using Genetic Algorithm (GA), Mutual Information (MI), and Information Gain (IG). BCI Competition III, Data Set I which has ECoG recordings motor imagery is used to evaluate the methodology. Results demonstrate that the feature selection improves the classification accuracy.