Minimizing intra class variations in multi-class common spatial patterns for motor imagery EEG signals

Brain Computer Interfaces have come a long way and have found practical use today. One of its major applications is controlling prosthetic devices for severely motor disabled patients. The Common Spatial Patterns is one the most efficient algorithms for extracting discriminative patterns for motor imagery tasks. The algorithm projects the data into a space where the discrimination between the classes is maximized. This algorithm is best suited for a two class paradigm. However, efficient algorithms for jointly approximately diagonalizing a set of matrices have made it possible to extend it to multi class problems. In this paper we use the QDIAG algorithm to approximately diagonalize variance matrices for more than two classes and our contribution to this is the optimization of the obtained filters so as to reduce the inter class dissimilarities for better discrimination. The results obtained show that this algorithm achieves better classification accuracies than other multi class feature extraction techniques.