Multiple SVM based Brain Computer Interface for Cybathlon 2020
Brain Computer Interfaces (BCI) are integrated software and hardware systems which record the bio-electrical signals of the brain and provide output for external devices. Our research team aims to create an EEG based BCI system, which will be challenged by other devices of competing research groups in May of 2020 at the international Cybathlon competition (https://cybathlon.ethz.ch/). In our initial paradigm, four motor movements and resting activity was imagined by the subjects, overall five imaginary classes. Our aim was to classify these command signals with an accuracy level, that allows an avatar to move continuously on the virtual race track of the competition. Feature extraction was based on the FFT power of each EEG channel. For classification method Support Vector Machine was selected and tested in two different strategies. In the first strategy the average activity of well known EEG bands (delta, alpha, beta) were used as features. With this method 28-34% accuracy level was achieved for five classes with single SVM approach. In the second strategy the average FFT power was considered in 2Hz bins of 14 bands. These features were trained in separate SVMs and the final classification was generated as maximum vote agreement of the individual SVMs. By the second method the level of classification accuracy increased to 38-44%. Multiple SVM classification could perform significantly better than individual one, therefore we suggest to move EEG signal classification towards ensemble SVM methods in order to achieve better results for BCI applications.