Detection of Event-Related Spectral Changes in Electrocorticograms

Abstract

The University of Michigan Direct Brain Interface (UM-DBI) project seeks to detect voluntarily produced electrocortical activity (ECoG) related to actual or imagined movements in humans as the basis for a DBI. In past work we have used cross-correlation based template matching (CCTM) as the method for detecting event-related potentials (ERPs). That approach ignores event-related spectral changes in the ECoG signal. This paper discusses model-based signal detection methods that exploit event-related spectral changes. In particular we propose a quadratic detector based on a two-class hypothesis test with different covariances for the two classes. The covariance matrices are generated by fitting autoregressive (AR) models to training data. Preliminary results show that the quadratic detector yields more channels with good detection performance than the CCTM method, particularly when we impose constraints on detection delay