Classification of Motor Imagery Task by Using Novel Ensemble Pruning Approach

Abstract

Brain-computer interface (BCI) assists communication for the disabled and handicapped. It is usually electroencephalogram (EEG) based and uses motor imagery (MI) in its operation. EEG signals are known for being nonstationary and are sensitive to artifacts from various sources such as the physical and mental state of the patient, their mood, their posture, and any external noise or distractions, etc. Processing of this data directly affects the classification accuracy, making it a critical step in any BCI system. Ensemble learning has been used for many kinds of BCI classification applications including MI and P300 event related potential, which has been proven to be robust. The purpose of this paper is to generate an algorithm that uses ensemble pruning method for EEG classification evoked by an MI task. In order to achieve this, we extracted the features of an EEG dataset and trained a range of support vector machines to make a diverse ensemble of classifiers. This ensemble is then pruned by using a novel optimization model by a difference of convex algorithm, which has not been used on EEG data before.