Offline detection of P300 in BCI speller systems

Abstract

The paper presents a framework for offline analysis of P300 speller system using seeded k-means based ensemble SVM. Due to the use of smalldatasets for the training of classifier, the performance deteriorates. The Proposed framework emphases on semi-supervised clustering approach for training the SVM classifier with large amount of data. The normalized mutual information (NMI) has used for cluster validation that gives maximum 88 clusters on 10 fold cross-validation dataset with NMI approx equals to 1. The proposed framework has applied to the EEG data acquired from two subjects and provided by the Wadsworth center for brain-computer interface (BCI) competition III. The experimental results show the increase in SNR value and obtain better accuracy results than linear, polynomial or rbf kernel SVM.