Improving Feature Selection in Riemannian Tangent Space: a New Approach for MI-BCI Detection
Brain–Computer Interfaces (BCIs) provide a non-muscular channel to control external devices using only brain activity. In particular, Motor Imagery BCIs (MI-BCIs) can be used to decode the imagination of certain movements for controlling rehabilitation technologies. Although algorithms based on Common Spatial Patterns (CSP) are widely used in MI-BCIs, they are not robust enough to data changes. Recently, Riemannian classifiers based on tangent space projection have been proposed as a promising approach for MI detection. These projections can be used as high dimensional feature vectors and then be classified with traditional machine learning methods. In this work, we tackle the high dimensionality of the tangent space by employing two feature selection methods previous to Linear Discriminant Analysis (LDA) classification: Stepwise LDA (SWLDA) and Least Absolute Shrinkage and Selection Operator (LASSO). The two proposed methods are compared with both, the traditional CSP framework and a simple classifier based on geometric distance between covariances matrices on the Riemannian space. The method based on LASSO feature selection yields the best performance for three real MI-BCI databases in a cross-validation scenario. For this approach, enhancements over CSP accuracy of up to 3.7% were found. For SWLDA, notable classification improvements were observed for specific subjects. These results clearly evidence that by an appropriate selection of features on the Riemannian tangent space MI detection can be improved, showing that these techniques are quite promising in the context of BCIs.