Small Sample Motor Imagery Classification Using Regularized Riemannian Features

Abstract

Motor imagery-based electroencephalogram brain-computer interface (BCI) performance suffers from huge variations within and across subjects. This is due to different spatial and temporal characteristics among the subjects. To address these variabilities, a large number of labeled subject specific training trials are collected to calibrate systems for new subjects. This results in long calibration time that limits the BCI usage in practice. One major challenge in the development of brain computer interface is to reduce calibration time or completely eliminate it. The existing approaches rise up to this challenge by incorporating inter-subject and intra-subject variations through covariance matrices from other subject's training trials. Current approaches use covariance matrices but do not consider the geometry of the covariance matrices, which lies in the space of symmetric positive definite matrices. This inevitably limits their performance. We focus on reducing calibration time by introducing a Riemannian approach. However, in Riemannian approach the performance degrades in small sample scenario as the dimensionality of covariance matrices is large in comparison to the number of trials. To overcome this limitation, we proposed a new framework that transforms covariance matrices into a lower dimension through spatial filter regularized by data from other subjects. The efficacy of the proposed approach was validated on the small sample scenario dataset IVa from BCI competition III. The proposed approach has achieved 87.21% mean accuracy and 0.74 mean kappa on dataset IVa. The proposed method outperforms the conventional method and other existing studies on dataset IVa. To ensure the robustness of the proposed method we evaluated on dataset IIIa from BCI competition III and dataset IIa from BCI competition IV. The proposed method has achieved mean accuracy 90.93% and 80.98% on dataset IIIa and dataset IIa, respectively.