33% Classification Accuracy Improvement in a Motor Imagery Brain Computer Interface

Abstract

A right-hand motor imagery based brain-computer interface is proposed in this work. Such a system requires the identification of different brain states and their classification. Brain signals recorded by electroencephalography are naturally contaminated by various noises and interferences. Ocular artifact removal is performed by implementing an auto-matic method “Kmeans-ICA” which does not require a reference channel. This method starts by decomposing EEG signals into Independent Components; artefactual ones are then identified using Kmeans clustering, a non-supervised machine learning technique. After signal preprocessing, a Brain computer interface system is implemented; physiologically interpretable features extracting the wavelet-coherence, the wavelet-phase locking value and band power are computed and introduced into a statistical test to check for a significant difference between relaxed and motor imagery states. Features which pass the test are conserved and used for classification. Leave One Out Cross Validation is performed to evaluate the performance of the classifier. Two types of classifiers are compared: a Linear Discriminant Analysis and a Support Vector Machine. Using a Linear Discriminant Analysis, classification accuracy improved from 66% to 88.10% after ocular artifacts removal using Kmeans-ICA. The proposed methodology outperformed state of art feature extraction methods, namely, the mu rhythm band power.