A Logistic Regression Based Framework for Spatio-Temporal Feature Representation and Classification of Single-Trial EEG

The classic motor imagery EEG signal analysis pipelines are implemented by two separate supervised stages (normally CSP+FLDA), note that the optimal solution is difficult to guarantee. Moreover, CSP only utilizes the spatial information os EEG signal, while neglecting the underlying temporal information. In this work, an alternative approach to CSP+FLDA is proposed (named LRSTC), in which only a single supervised learning stage is needed. By LRSTC, the feature extraction and classification can be tackled conveniently under a regularized empirical risk minimization problem. The input signal the whitened spatial covariance matrices, and we use a linear model to simultaneously learn the spatio-temporal filters and the weights of classifier. To address the potential over-fitting issue, an nuclear norm is added in our objective function as the regularization term. One motor imagery EEG data set from past BCI competitions is used to evaluate the performance of our algorithm. Compared with the CSP+FLDA, FBCSP+FLDA, the algorithm proposed by Tomioka (termed as LRC in this paper), and CSSSP+FLDA, our algorithm shows significant classification performance except for FBCSP+FLDA.