Under-sampling and Classification of P300 Single-Trials using Self-Organized Maps and Deep Neural Networks for a Speller BCI

A Brain-Computer Interface (BCI) allows its userto control machines or other devices by translating its brainactivity and using it as commands. This kind of technologyhas as potential users people with motor disabilities since itwould allow them to interact with their environment withoutusing their peripheral nerves, helping them to regain their lostautonomy. One of the most successful BCI applications is theP300-based Speller. Its operation depends entirely on its capacityto identify and discriminate the presence of the P300 potentialsfrom electroencephalographic (EEG) signals. For the system to dothis correctly, it is necessary to choose an adequate classifier andtrain it with a balanced data-set. However, due to the use of anoddball paradigm to elicit the P300 potential, only unbalanceddata-sets can be obtained. This paper focuses on the trainingstage of two classifiers, a deep feedforward network (DFN) anda deep belief network (DBN), to be used in a P300-based BCI. Thedata-sets obtained from healthy subjects and post-stroke victimswere pre-processed and then balanced using a Self-OrganizingMaps-based under-sampling approach prior training looking toincrease the accuracy of the classifiers. We compared the resultswith our previous works and observed an increase of 7% inclassification accuracy for the most critical subject. The DFNachieved a maximum classification accuracy of 93.29% for apost-stroke subject and 93.60% for a healthy one.