Improving Speller BCI performance using a cluster-based under-sampling method

A Brain-Computer Interface (BCI) allows its user to interact with a computer or other machines by only using their brain activity. People with motor disabilities are potential users of this technology since it could allow them to interact with their surroundings without using their peripheral nerves, helping them regain their lost autonomy. The P300 Speller is one of the most popular BCI applications. Its performance depends on its classifier's capacity to identify and discriminate the presence of the P300 potentials from electroencephalographic (EEG) signals. For the classifier to do this correctly, it is necessary to train it with a balanced data-set. However, as the P300 is usually elicited with an oddball paradigm, only unbalanced distributions can be obtained. This paper applies an under-sampling method based on Self-Organizing Maps (SOMs) on P300 EEG signals looking to increase the classifier's accuracy. Two classifying models, a deep feedforward network (DFN) and a deep belief network (DBN), are tested with data-sets obtained from healthy subjects and post-stroke victims. We compared the results with our previous works and observed an increase of 7% in classification accuracy for our most critical subject. The DBN achieved a maximum classification accuracy of 95.53% and 94.93% for a healthy and post-stroke subject, while the DFN, 96.25% and 93.75%.