An empirical study to evaluate feature extraction approaches CSP, TSM, and CSP-TSM on a MI-BCI under distraction

Common spatial patterns (CSP) and tangent space mapping (TSM) are frequently used approaches for Motor Imagery Brain-Computer Interface (MI-BCI). These two methods are used in the feature extraction block to transform the EEG time series into a set of features to facilitate the classification learning process. CSP is based on spatial filtering, while TSM is based on covariance matrix estimation and Riemannian geometry framework, a third approach called CSP-TSM can be used by combining CSP and TSM. In this paper, we compare experimentally the classification accuracy and the computational time of these three approaches (CSP, TSM, and CSP-TSM) on a MI-BCI dataset acquired under five types of distractions that simulates a pseudo-realistic environment. The obtained results help us to explore the pros and cons of using each approach in a MI-BCI operated out-of-lab environment.