How many people can control a motor imagery based BCI using common spatial patterns

EEG based Brain-Computer Interfaces (BCIs) often use evoked potentials (P300), steady state visual evoked potentials (SSVEP) or motor imagery (MI) for control strategies. This study investigated maximum and mean accuracy of a MI based BCI using Common Spatial Patterns (CSP). Twenty healthy people participated in the study and were equipped with 64 active EEG electrodes. They performed a training paradigm with 160 trials by imagining either left or right hand movement to set up a subject specific CSP filter to spatially filter the EEG data. Following that, two real-time runs with 80 trials were performed, which provided feedback to the subject. The real-time accuracy was then calculated for every subject, and finally a grand average accuracy of 80.7% was reached for the 20 subjects. One person reached a perfect classification result of 100%, 30% performed above 90% and one was below 59%. The results show that most people can use a MI based BCI after a brief training time if CSPs with 64 active electrodes are used. The method of CSP yields clearly better classification results compared to a bandpower approach. While more electrodes are needed for classification, this is less of a disadvantage with modern active electrodes.