Neuro-evolutionary approach for optimal selection of EEG channels in motor imagery based BCI application

摘要

The selection of relevant channels that produces optimal subset of electroencephalogram (EEG) features is of prime importance for i) reducing computational complexity, ii) reducing overfitting, and iii) eliminating inconveniences in users during clinical application. In the analysis of EEG signals acquired directly from the scalp, feature extraction and classification techniques play a key role in motor imagery based brain-computer interface (BCI) system. Typically, the high-density EEG system used in this process is composed of 64 or more channels for EEG signal recordings. However, in practical applications, fewer channels are required for the system set up. Thus, for the current study, we aim to determine the optimal number of channels for the BCI system. We propose a channel selection strategy known as neuro-evolutionary algorithm, which is based on modified particle swarm optimization (MPSO), coupled with feature extraction by common spatial pattern. To validate the proposed method, we performed experiments using 64 channel EEG recordings from four trans-humeral amputees. The acquired signals from five motor imagery tasks were fed into a wrapper based neuro-evolutionary framework after the extraction stage. The performance results of the MPSO, when compared with other models like ant colony, genetic algorithm, particle swarm optimization and simulated annealing, demonstrate the superiority of MPSO as the best performance with the most significant results. Further validation with electrocorticogram (ECoG) dataset IV of the BCI competition showed that the proposed model outperformed other commonly used methods by significantly decreasing the error rate and the number of EEG channels.