Classification of Brain States using Subject-Specific Trained Classifiers

Abstract

One of the key components to develop a usable Electroencephalography (EEG) based Brain Computer interface (BCI) is the efficient classification of EEG patterns using Machine Learning classifiers. This paper presents a comparison of Linear Discriminant Analysis (LDA), Naïve Bayes and Decision Tree classifiers by applying to the EEG data. The classifiers are applied on BCI Competition III: Dataset V that consists of three cognitive tasks, namely, right and left hand imagery movement and the imagination of any word starting from a given random letter. The BCI experiments for this data have been performed with three subjects. For subjects 1 and 2, the Naïve Bayes classifier provides best results while for subject 3 the maximum accuracy is achieved from LDA classifier. In order to improve the accuracy further, it has been proposed to apply combination of classifiers based on Multiple/Ensemble Classifier System concept on data for single subject with different sessions of data recording. By combining the classifiers LDA and Decision Tree, maximum accuracies of 81%, 70% and 56% for subjects 1, 2 & 3 respectively have been achieved that are comparable with the accuracies achieved by the winner of the competition. It is concluded that instead of employing single classifier, the approach for using combination of classifiers significantly improves the performance of a BCI system.