Brain imaging and support vector machines for brain computer interface

Abstract

Signal subspace correlation methods are used to derive EEG features for a brain computer interface (BCI) system. The "multiple signal classification" (MUSIC) algorithm was applied to scan a single dipole model through a grid confined to a three dimensional head model. The projection onto an estimated signal subspace was then computed to extract relevant features that were provided to a classifier whose aim was to determine the request conveyed by the user. Two classifiers, the multilayer perceptron (MLP) and the support vector machines (SVM) were tested and compared. The use of SVM with features extracted from signal subspace correlation yielded an error rate of 17% on a reference database suggesting that the proposed BCI system shows better results than the known state of the art systems.