Single-trial P300 classification using deep belief networks for a BCI system

A brain-computer interface (BCI) aims to provide its users with the capability to interact with machines only through its brain activity. There is a special interest in developing BCIs targeted at people with mild or severe motor disabilities since this kind of technology would improve their lifestyles. The Speller is a BCI application that uses the P300 waveform to essentially allow its user to communicate without using its peripheral nerves. This paper focuses on the classification of the P300 waveform from single-trials obtained through EEG using deep belief networks (DBNs). This deep learning algorithm can identify relevant features automatically from the subject’s data, making its training requiring less pre-processing stages. The network was tested using signals recorded from healthy subjects and post-stroke victims. The highest accuracy achieved was of 91.6% for a healthy subject and 88.1% for a post-stroke victim.