Reconstruction of cortical sources activities for online classification of electroencephalographic signals

We compare the results given by different methods to reconstruct cortical sources activity in order to classify EEG in real time. Two motor imagery experiments were performed. The aim was to retrieve from 1-second windows of signal which motor imagery task the subjects were performing. The use of cortical activity reconstruction was compared to Laplacian filtering, which is often used in BCI. A recursive algorithm using Student's t-test was used to select relevant cortical sources. The Beamformer method led to an improvement of the classification for the first experiment, which included six motor imagery tasks. The weighted Minimum-Norm method required the use of a specific head model, extracted from the subject's MRI, to improve the classification. It then gave the best results on the second experiment, achieving a classification rate of 77% compared to 71% for direct use of electrode data and 75% for Laplacian filtering and Beamformer.