Bayesian classification of index finger movements by analysis of MEG and EEG data

Abstract

Abstract A Bayesian classifier was developed for decoding finger movements by analysis of MEG data. Subjects were monitored by a 151-channel MEG system while making “flicking” motions (up, down, left, right) of the right-hand index finger. The SAM beamformer method was used to spatially localize the brain activity. A classifier was constructed based on signals from 30 to 200 discriminating locations selected from the 1-mm grid. When applied to the test data, 4-way classification rates in the range 50–70% were observed (chance = 25%), with information rates of 0.25–0.7 bits per classification. In several cases simultaneous EEG recordings were made. By calculating regression of SAM signals from discriminating locations on EEG training data, it was possible to accomplish the “informing” of the EEG by MEG. It was shown that even when classification using physical EEG channels failed, informed EEG yielded 40–45% classification rates.