Feature Selection methods applied to Motor Imagery task classification

摘要

Brain Computer Interfaces allow the interaction between a person and their environment using signals extracted directly from the brain. One of the most common non-invasive methods of brain signal acquisition is the electroencephalography (EEG). An EEG based BCI system requires the processing and translation of the EEG signal into significant features that could be converted into commands for the external agent. This system generally involves four steps: preprocessing, feature extraction, feature selection and classification. To guarantee a good classification performance, the application of a convenient feature selection algorithm has a great importance. However, in the literature the information is scarce about a comparison of the different methods, making very difficult the selection of an useful method for a BCI system. In this paper, six feature selection (all separately applied to EEG data) methods: CFS, Consistency, ReliefF, mRmR, C4.5 and Genetic Algorithm are evaluated and compared for Motor Imagery task classification. The validation is made by the performance of five classifiers commonly employed on EEG data: PNN, RBF, SVM, LDA and k-NN. The best result (93,71% of correct classifications) was obtained by the Genetic Algorithm combined with the LDA classifier.