Detection and classification of tongue movements from single-trial EEG

AIM: To detect and classify tongue movements from single trial electroencephalography (EEG), so that it can be used as a reliable control signal in a brain computer interface (BCI). Method: Thirteen subjects, all BCI-naïve, performed four different tongue movements (up, down, left and right), which was detected against an idle state using a common spatial pattern filter with a linear discriminant analysis classifier. Furthermore, the movement types were classified in a one-versus all classification scheme. Results: On average, 72-76% of the movements were detected correctly against the idle state. When all movement types were pooled and detected against the idle state, an accuracy of 80% was obtained. A closer investigation showed that the system correctly detected up to 83% of the executed movements, but had a false positive rate of 13%. The movements were classified with an accuracy of 43%. This was increased to 55% when only left, right and up movements were considered. When only left and right movements where considered they were classified with an average accuracy of 71%. Conclusion: Decoding of tongue movements from the EEG can be used as a reliable control state switch in a BCI and is possible to classify the different movements above chance level. Significance: Residual tongue movements, which is not lost after a spinal cord injury, can be used as a reliable control state switch and it is possibly to detect at least four different movement types.