Channel Selection Improves MEG-Based Brain-Computer Interface

Abstract

This study investigates the effect of channel selection on the performance of a Magnetoencephalography (MEG)-based brain-computer interface (BCI) system in terms of classification accuracy (CA). Although many efforts are currently being undertaken to develop BCI using MEG, the major concern still is low accuracy. MEG systems involve data recording from a large number of channels which may provide a better spatio-temporal resolution for assessing brain patterns, however, a large numbers of channels result in a large number of features, which further make feature learning a challenging task. In this study, we evaluated the performance of two state-of-the-art channel selection methods, i.e. class-correlation (CC) and ReliefF (RF) across six binary classification tasks with a MEG dataset of 15 healthy participants. Both CC and RF methods provided a statistically significant increase in the CA (range: 20.91 - 24.22%) compared to baseline (i.e. using 204 channels) with bandpower features from the alpha (8-12 Hz) and beta frequency bands (13-30 Hz). Moreover, both methods reduce the optimum number of channels significantly (from 204 to the range of 1-22). Reducing the number of features can significantly reduce the computational cost and increase the chances of numerical stability which are key considerations in neurofeedback (online) applications.