Shrinkage estimator based regularization for EEG motor imagery classification

Electroencephalography (EEG) based motor imagery Brain-Computer Interface (MI-BCI) paradigm is used to communicate with external device by people who lost peripheral nerve control, or perform neuro-rehabilitation for stroke patients. BCI systems based on motor imagery often employ feature extraction algorithms based on Common Spatial Patterns (CSP). CSP is capable of discriminating two classes, but it is sensitive to outliers and noisy trials. Therefore, regularisation is often deployed to improve the robustness and accuracy of CSP estimation. In this paper, a novel regularisation approach based on shrinkage estimation is presented in order to handle small sample problem and retain subject-specific discriminative features. In this method, an analytical solution for shrinkage estimation is provided, which not only is computationally tractable, but also overcomes the heuristic approach of traditional cross validation based parameter tuning. We applied the proposed regularisation to Filter-Bank Common Spatial Pattern (FBCSP). The proposed method is evaluated on two publicly available datasets, namely Wadsworth Physiobank Dataset and BCI Competition IV Dataset 2a. The results show that Shrinkage Regularized Filter Bank CSP (SR-FBCSP) outperforms FBCSP in classifying left vs right hand motor imagery.