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

• Published in: Conference proceedings - IEEE International Conference on Systems, Man, and Cybernetics pp. 2972 - 2978
• Main Authors: Cortez, Sergio A., Flores, Christian, Andreu-Perez, Javier
• Format: Conference Proceeding
• Language: English
• Published: IEEE 11.10.2020
• Subjects: brain-computer interface; Brain-computer interfaces; Computer architecture; EEG; Electric potential; Electroencephalography; Medical conditions; Neural networks; post-stroke; self-organizing maps; Training
• ISSN: 2577-1655
• DOI: 10.1109/SMC42975.2020.9283178

A Brain-Computer Interface (BCI) allows its user to control machines or other devices by translating its brain activity and using it as commands. This kind of technology has as potential users people with motor disabilities since it would allow them to interact with their environment without using their peripheral nerves, helping them to regain their lost autonomy. One of the most successful BCI applications is the P300-based Speller. Its operation depends entirely on its capacity to identify and discriminate the presence of the P300 potentials from electroencephalographic (EEG) signals. For the system to do this correctly, it is necessary to choose an adequate classifier and train it with a balanced data-set. However, due to the use of an oddball paradigm to elicit the P300 potential, only unbalanced data-sets can be obtained. This paper focuses on the training stage of two classifiers, a deep feedforward network (DFN) and a deep belief network (DBN), to be used in a P300-based BCI. The data-sets obtained from healthy subjects and post-stroke victims were pre-processed and then balanced using a Self-Organizing Maps-based under-sampling approach prior training looking to increase the accuracy of the classifiers. We compared the results with our previous works and observed an increase of 7% in classification accuracy for the most critical subject. The DFN achieved a maximum classification accuracy of 93.29% for a post-stroke subject and 93.60% for a healthy one.