An ensemble classification approach to motor-imagery brain state discrimination problem

This paper focuses on the use of ensemble classifiers on motor imagery data to distinguish brain states. Raw EEG signal is filtered and represented separately in terms of following features: Band power (BP), wavelet based energy-entropy (Engent) and feature extracted with adaptive autoregressive (AAR) model. We tested the classifiers using both hold-out testing (termed Experiment-I) and 10-fold cross-validation with stratified sampling (called Experiment II). We observe from our empirical study that the ensemble classifier particularly the subspace variant outperforms others in terms of classification accuracies in both experiment-I and II. Features extracted with AAR and energy-entropy techniques provide most consistence performance for experiment-I and II respectively.