Online processing for motor imagery-based brain-computer interfaces relying on EEG

This manuscript reports a comparison among three possible strategies for online processing of electroencephalo-graphic signals, in terms of their impact on the online classification accuracy. The comparison was carried out in the framework of brain-computer interfaces based on motor imagery. Filter bank common spatial pattern was exploited as a standard feature extraction technique along with a support vector machine for classification of the brain signals. This machine learning-based algorithm was trained offline and evaluated on independent evaluation data by means of the online processing strategies. Benchmark dataset were used, so that the online processing performance was compared to reference offline performances compatible with literature (at least 74% classification accuracy). Results suggest that it is convenient to use the bigger part of the imagery period in training the algorithm prior to online classification accuracy. Moreover, using an enlarging window for evaluation appeared to be the best strategy to remain close to reference mean accuracy.