Impact of NSGA-II objectives on EEG feature selection related to motor imagery

The selection of ElectroEncephaloGram (EEG) features with functional relevance to Motor Imagery (MI) is a crucial task for successful outcome in Brain-Computer Interface (BCI)-based motor rehabilitation. Individual EEG patterns during MI requires subject-dependent feature selection, which is an arduous task due to the complexity and large number of features. One solution is to use metaheuristics, e.g. Genetic Algorithm (GA), to avoid an exhaustive search which is impractical. In this work, one of the most widely used GA, NSGA-II, is used with an hierarchical individual representation to facilitate the exclusion of EEG channels irrelevant for MI. In essence, the performance of different objectives in NSGA-II was evaluated on a previously recorded MI EEG data set. Empirical results show that k-Nearest Neighbors (k-NN) combined with Pearson's Correlation (PCFS) as objective functions yielded higher classification accuracy as compared to the other objective-combinations (73% vs. 69%). Linear Discriminant Analysis (LDA) combined with Feature Reduction (FR) as objective functions maximized the reduction of features (99.6%) but reduced classification performance (65.6%). All classifier objectives combined with PCFS selected similar features in accordance with expected activity patterns during MI. In conclusion, PCFS and a classifier as objective functions constitutes a good trade-off solution for MI data.