Multichannel Classification of Single EEG Trials with Independent Component Analysis

We have previously shown that classification of single-trial electroencephalographic (EEG) recordings is improved by the use of either a niultichannel classifier or the best independent component over a single channel classifier. In this paper, we introduce a classifier that makes explicit use of multiple independent components. Two models are compared. The first ("direct") model uses independent components as time-series inputs, while the second ("indirect") model remixes the components back to the signal space. The direct model resulted in significantly improved classification rates when applied to two experiments using both monopolar and bipolar settings.