Supervised adaptive downsampling for P300-based brain computer interface

To realize brain computer interface, a recording electroencephalogram (EEG) and determining whether or not P300 is evoked by the presented stimulus have become increasingly important. Using the machine learning method for this classification is effective, but constructing feature vectors with all data points might result in very high-dimensional data. Because such redundant features are undesirable from the viewpoint of computation and classification performance, EEG has been downsampled in several studies. In the present study, we propose a new downsampling method aiming at the improvement of P300 classification accuracy. In particular, each single trial EEG is segmented at non-uniform intervals and then averaged in each segment. The segmentation is decided in such a way that the degree of separating two classes from training data is increased by applying a time series segmentation algorithm. Our experiment using the BCI Competition III P300 Speller paradigm data set demonstrated that our method resulted in higher accuracy than traditional downsampling methods.