A BCI system based on orthogonalized EEG data and multiple multilayer neural networks in parallel form

Abstract

A BCI system, using orthogonalized EEG data sets and multiple multilayer neural networks (MLNNs) in a parallel form, is proposed. In order to emphasize feature of multi-channel EEG data, Gram-Schmidt orthogonalization has been applied. Since there are many channel orders to be orthogonalized, many kinds of orthogonalized data sets can be generated for the same EEG data set by changing the channel order. These data sets have different features. In the proposed method, different channel orders are assigned to the multiple MLNNs in a training phase and in a classification process. A good solution can be searched for by changing the channel orders within a small number of trials. By using EEG data for five mental tasks, a correct classification rate is increased from 88% to 92%, and an error classification rate is decreased from 4% to 0%. © 2010 Springer-Verlag Berlin Heidelberg.