A deep learning approach to single-trial classification for P300 spellers

Brain-Computer Interface (BCI) is one of the most promising fields nowadays in hope to assist individuals with cognitive or sensory-motor disabilities. It facilitates human interaction with modern technologies and tools through brain activity. This could grant some communication capabilities back to severely disabled individuals. The P300 speller is one of the most successful BCI communication applications. A typical P300 speller is presented to the user as a 6-by-6 matrix whose cells contain the essential letters, digits, and characters that are randomly flashed. The usage of this application typically takes a significant amount of time due to the need to perform multiple trials to accurately recognize the target cell. We introduce a deep neural network approach to enhance the recognition of the target cell from a limited number of trials. We recorded the electroencephalography (EEG) signals from two subjects using Emotiv Epoc neuroheadset. A deep neural network consisting of multiple autoencoder layers and a softmax classification layer was trained with filtered EEG data. Compared to using Principal Component Analysis (PCA) and linear classifiers, the proposed model achieved an average increase in accuracy of 4.5% when applied to averaged 5 trials P300 data, and, more importantly, an increase of 8% when applied to single trial P300 data. The obtained results elucidate the significant effect of deep learning techniques in this paradigm.