GA methods for selecting the proper EEG individual spectral bands limits

Abstract

In a previous set of analyses and researches we have proved the strong relationship that exists between each particular subject and its corresponding particular set of imagery cognitive tasks (determined out of several proposed mental tasks); these individual sets of tasks were the ones on which the obtained classification performances were significantly superior than on the other possible combinations of tasks. Also, a remarkable aspect is that all these improvements in classification were achieved for the same EEG features (namely, AR parameters) and the same processing and classification methods, that, during the entire study, were kept unmodified. In consequence, the act of finding, for each individual subject, the appropriate specific set of cognitive tasks should be considered of great importance in any brain computer interface (BCI) implementation. The present paper continues these researches and focuses on the necessity to find (as it has been already suggested in the literature), for each subject, that set of custom band power coefficients for which superior classification rates on the subject optimal set of cognitive tasks - previously determined - will be the highest. Based on some specific GA methods, implemented in order to find the subject appropriate frequency band parameters, and using a neural network structure for the classification, the final new obtained classification performances considerably improved with 4 to 6 percents.