Improving P300 Speller performance by means of optimization and machine learning

Brain-Computer Interfaces (BCIs) are systems allowing people tointeract with the environment bypassing the natural neuromuscular and hor-monal outputs of the peripheral nervous system (PNS). These interfaces recorda user's brain activity and translate it into control commands for external de-vices, thus providing the PNS with additional artificial outputs. In this frame-work, the BCIs based on the P300 Event-Related Potentials (ERP), whichrepresent the electrical responses recorded from the brain after specific eventsor stimuli, have proven to be particularly successful and robust. The presenceor the absence of a P300 evoked potential within the EEG features is deter-mined through a classification algorithm. Linear classifiers such as stepwiselinear discriminant analysis (SWLDA) and support vector machine (SVM)are the most used discriminant algorithms for ERPs' classification. Due to thelow signal-to-noise ratio of the EEG signals, multiple stimulation sequences(a.k.a. iterations) are carried out and then averaged before the signals be-ing classified. However, while augmenting the number of iterations improvesthe Signal-to-Noise Ratio (SNR), it also slows down the process. In the early studies, the number of iterations was fixed (no stopping environment), butrecently several early stopping strategies have been proposed in the literatureto dynamically interrupt the stimulation sequence when a certain criterion ismet in order to enhance the communication rate. In this work, we explore howto improve the classification performances in P300 based BCIs by combiningoptimization and machine learning. First, we propose a new decision functionthat aims at improving classification performances in terms of accuracy andInformation Transfer Rate both in a no stopping and early stopping environ-ment. Then, we propose a new SVM training problem that aims to facilitatethe target-detection process. Our approach proves to be effective on severalpublicly available datasets.