Feature extraction and classification of brain motor imagery task based on MVAR model

In this paper, MVAR (multivariate autoregressive) model method for extracting EEG features is presented. With MVAR model coefficient features, the discriminant analysis based on Mahalanobis distance is applied to realize classification of the left and right hand motor imagery tasks. By analyzing the data from BCI2003 competition provided by Graz University of technology, the satisfactory results are obtained with the highest classification accuracy reaching 88.57% and the maximum mutual information reaching 1.03 bit. To testify the validity of MVAR model method, as a contrast EEG feature extraction by AR model is discussed. From the three performances such as maximum classification accuracy, maximum SNR and maximum mutual information, the results by MVAR method are better than that by AR model method.