Feature Analysis for Motor Imagery EEG Signals with Different Classification Schemes

Abstract

A Brain-Computer Interface (BCI) is a communication system that decodes and transfers information directly from the brain to external devices. The electroencephalogram (EEG) technique is used to measure the electrical signals corresponding to commands occurring in the brain to control functions. The signals used for control applications in BCI are called Motor Imagery (MI) EEG signals. EEG signals are noisy, so it is important to use the right methods to recognize patterns correctly. This study examined the performances of different classification schemes to train networks using Ensemble Subspace Discriminant classifier. Also, the most efficient feature space was found using Neighborhood Component Analysis. The maximum average accuracy in classifying MI signals corresponding to right-direction and left-direction was 80.4% with a subject-specific classification scheme and 250 features.