A Study via Feature Selection on the Separability of Approximate Entropy for Brain-Computer Interfaces

The linear and nonlinear separability of approximate entropy feature for EEG-based brain-computer interfaces (BCI) are tested and compared to that of two alternative features, band power and reflection coefficients. Separability is analyzed using hybrid sequential forward floating search, in which a classifier: linear discriminate analysis (LDA) classifier or nonlinear support vector machine (SVM), and a separability index: Davies-Bouldin index (DBI) or mutual information (MI) based index, are jointly utilized to evaluate selected feature subsets. Results on BCI data demonstrate the separability of the approximate entropy feature to be comparable to that of the band power and reflection coefficients, although each feature has advantages on different situations.