Smooth bilinear classification of EEG.

The goal of this paper is to improve on single-trial classification of electro-encephalography (EEG) using linear methods. The paper proposes to combine the classification of the spatial distribution of activity with the classification of its temporal profile. The work is based on the idea that a current source in the brain has a reproducible temporal profile with a static spatial projection to the electrodes. This assumption reduces the parameter space of a linear classifier to a rank-one factorial space. The new model limits over-fitting due to the fewer number of parameters, and furthermore, it allows us to declare a prior belief of smoothness on the spatial and temporal profiles of the source. Our experiments show that the method is useful as a classifier with an area under the ROC curve of 0.93 having only 40 target trials available for training. Investigation of the trained classifier encourages us to belief that the method can also be useful as a tool to interpret the activity in the data at hand with respect to experimental events.