Investigation of the effect of channel selection algorithms on recognition performance in spelling systems from EEG signals

With several decades of history, BCI are being developed to enable individuals with such disorders to communicate with their environment. Speller systems are detected characters of the person focused in a matrix consisting of letters and numbers on a screen and convert them into writing through the application. The problem is large data and the associated processing costs are high. Reduce the number of channels used for reduce the amount of data and process cost. Optimization techniques come into play with metaheuristic methods in cases where which channel is more efficient cannot be designed with mathematical model. The study was conducted on the data called BCI III Competition Dataset II held in 2005. In this study, channels were selected by the genetic algorithm, binary particle swarm optimization and binary artificial bee colony. The selected channels were grouped according to the number of selections in the repetitions by 8 channels, 16 channels and average number of selections. The performance of the selected channels in the singular and ensemble structure was investigated by using SVM, LS-SVM and LDA. The classifier performance analysis was accuracy, sensitivity, precision, F score measurement. As a result of the study, it has been observed that channel reduction causes performance increase for all classifiers. Especially the data with low accuracy rate increases at a higher rate. After the channel selection, accuracy rates exceeding 97% were obtained. This metaheuristic methods are used in channel selection processes in BCI systems, decreasing the amount of data and the transaction cost.