Improved classification of motor imagery datasets for BCI by using approximate entropy and WOSF features

A Brain Computer Interfaces (BCI) system enables users to control devices by acquiring and processing brain activity. An important component of a BCI system is feature extraction, which is responsible for representing brain signals in terms of essential components called features. This paper presents a comparison of the following feature extraction techniques for BCI; Common Spatial Patterns (CSP), Wavelength Optimal Spatial Filter (WOSF) and Approximate Entropy. The motivation for this work is the non-availability of comparative studies on the mentioned feature extraction techniques in literature. Further, even though CSP has been a widely used feature extraction technique for motor-imagery based BCI systems, entropy-based features, such as approximate entropy, and WOSF are still being explored. We investigate the use of approximate entropy and WOSF for feature extraction in motor imagery datasets of BCI Competitions, and compare the results with those obtained using CSP. It was observed that both WOSF and Approximate Entropy provide a higher classification accuracy as compared to CSP.