Automated classification of non-motor mental task in electroencephalogram based brain-computer interface using multivariate autoregressive model in the intrinsic mode function domain

Abstract Objective In this paper, we proposed a new feature extraction approach based on the multivariate auto regressive model (MVAR) model of the sensitive intrinsic mode function (IMF) groups in the multivariate empirical mode decomposition (MEMD) domain. Approach We computed eigen values from the coefficient matrix of the MVAR model for classifying three different non-motor cognitive task in EEG based brain computer interface (BCI) system. In the first stage, the application of MEMD to multichannel EEG data gave rise to adaptive i.e. data driven decomposition of the multivariate time series data into a large number of IMF groups. In the second stage, the sensitive IMF groups were selected according to their task correlation factor. MVAR model of order six was developed from the five sensitive IMF groups and finally the eigen values of the correlation matrix derived from the coefficient matrix was employed for forming the feature vectors. At the last stage, the extracted feature vectors were fed to a Least Squares Support Vector Machine (LS-SVM) classifier for automatic classification of mental task EEG signals. We tested our approach on the mental task EEG data sets of three subjects. Main result We achieved highest value of average classification accuracy of 94.43% for binary classification of the first pair of mental task i.e baseline and mental arithmetic task using polynomial kernel and 91.65% for the second pair i.e mental arithmetic and mental letter composing task using radial basis function (RBF) with ten fold cross validation. We achieved highest value of average classification accuracy of 77.77% for three class classification employing One Vs One scheme of multiclass SVM classifier. Significance The performance of the binary classifier was evaluated on various parameters such as accuracy, specificity and sensitivity. The encouraging results show the potential of the proposed approach for classifying any non linear and non-stationary signals.