Optimal Features for Cross Subject Classification of Imagined Left and Right Fist Movements using EEG Signals

Classifying limb movements using neural activity is an essential task in the Brain-Computer Interface. A cross-subject classification scheme based on the EEG signals for categorization of the imagined left versus right fist movements is presented. A variation of the Particle Swarm Optimization (PSO) is implemented for feature and channel selection. The proposed methodology is tested on an EEG Motor Movement/Imagery dataset. Based on the selected feature sets including spectral features, power spectral density, phase locked values, and wavelet coefficient based statistics, the K nearest neighbor classifier outperformed existing frameworks based only on single modality feature sets. The proposed method has achieved an average accuracy of 70.2% for the EEG data from 100 subjects and 74.7% for the data from 50 subjects.