Automatic adaptation to the beta rebound after brisk movement imagery in a brain-computer interface

We simulate how a two-class brain-computer interface automatically adapts to post-movement imagery bursts of beta band activity (beta rebound) measured in the electroencephalogram at Cz. We used data from 20 healthy, novice volunteers. By combining an adaptive BCI approach with beta rebound features we hypothesize to attain better performance for more users, higher usability and lower setup time than with previous approaches. Our simulation processed data trialwise: The adaptive BCI continuously performed trial based outlier rejection, auto-calibrated a linear classifier after ten trials per class, and re-calibrated at every five trials per class. We simulated online performance by always applying the most recent classifier to newly processed trials. We found a high average peak accuracy of 76.4 ± 10.6% over the participants. The present system performs equally well as a comparable state-of-the-art, low-scale co-adaptive BCI, but requires less user effort, a lower number of sensors and lower system complexity. The system also well complements existing beta rebound based BCI systems: In comparison to even simpler approaches it tends to work for more users. Compared to an approach that used motor execution to setup a classifier, the present system allows for faster, more intuitive and more effective calibration. We consider the encouraging results from this simulation an important step towards online operation.