Noise benefits in the array of brain-computer interface classification systems

Abstract

Abstract This paper shows that noise can improve the accuracy of brain-computer interface (BCI) systems. Additive Gaussian noise can benefit arrays of ensemble support vector machines (ESVMs) that classify P300 or motor imagery (MI) activities in electroencephalogram (EEG) signals. We show these noise benefits in 64-channel EEG signals from the BCI Competitions II dataset IIb and BCI Competitions III dataset II for the P300 speller paradigm and in 3-channel EEG signals from the BCI Competitions II dataset III and BCI Competitions III dataset IIIa for MI classification systems. We also show that noise can improve the accuracy of EEG classifications based on restricted channel positions in commercial recording systems, such as the 14-channel Emotiv Epoc headset. The experimental results show that noise can provide classifiers with higher accuracy and can reduce the data collection time for P300 classification. The results also show that training ESVMs with a concatenated original dataset and noise-added datasets can improve MI classification. Noise can improve the accuracy of P300 classification for both intra-subject and inter-subject classification systems for multiple users. Addition of noise can significantly affect the parameters of polynomial kernel functions and the number of support vectors of the SVM. This leads to an expansion of the margin between two parallel hyperplanes that eventually improve the classification accuracy. Particle swarm optimization (PSO) can be used to search for the optimal noise intensity.