MINING MULTICHANNEL EEG FOR ITS INFORMATION-CONTENT - AN ANN-BASED METHOD FOR A BRAIN-COMPUTER INTERFACE

We have studied 56-channel electroencephalograms (EEG) from three subj ects who planned and performed three kinds of movements, left and righ t index finger, and right foot movement. Using autoregressive modeling of EEG time series and artificial neural nets (ANN), we have develope d a classifier that can tell which movement is performed from a segmen t of the EEG signal from a single trial. The classifier's rate of reco gnition of EEGs not seen before was 92-99% on the basis of a 1 s segme nt per trial. The recognition rate provides a pragmatic measure of the information content of the EEG signal. This high recognition rate mak es the classifier suitable for a so-called 'Brain-Computer Interface', a system that allows one to control a computer, or another device, wi th ones brain waves. Our classifier Laplace filters the EEG spatially, but makes use of its entire frequency range, and automatically locate s regions of relevant activity on the skull. (C) 1998 Elsevier Science Ltd. All rights reserved.