Enhanced Generalization Performance for Motor Imagery BCI via Sensory Stimulation

Abstract

In this work, sensory stimulation (SS) is introduced for motor imagery (MI) decoding. Subjects are instructed to focus on sensation when the left or right wrist is tactile stimulated during the SS block, and they are required to perform left or right-hand imaginary movements during the MI block. The experiment paradigm involves three blocks, namely Pre-MI block, SS block, and Post-MI block. Each block contains 80 trials, and the Post-MI block is served as an evaluation block. Three calibration protocols are developed for the comparison of generalization performance. Concretely, MI calibration using trials from the Pre-MI block, SS calibration using trials from the SS block, and Mixed calibration using the first two blocks together. Four BCI naïve subjects are recruited in the study. The calibration is conducted by using common spatial pattern (CSP) features with the linear support vector machine (SVM). The experiment results have shown that SS calibration can be used for MI decoding. Meanwhile, Mixed calibration yields better generalization performance compared to MI calibration. The tactile sensation-induced oscillatory dynamics are similar to the one induced by MI. The proposed Mixed calibration is a promising calibration protocol for MI decoding.