The University of British Columbia

A technique which utilizes outlier detection theory to extract event-related information from single-trial EEG is described. The technique uses generalized maximum-likelihood parameter estimation to provide robust autoregressive models of the EEG signal. These models are then used in a robust signal estimator to produce a cleaned version of the signal. The outlier content, the difference between the original and the cleaned signal, is then processed to yield event-related information. The technique was applied to EEG signals recorded during a voluntary skilled thumb movement. The results showed that consistent motor-related information can be obtained on a single-trial basis.