Enhance decoding of lower limb motor imagery‐electroencephalography patterns by Riemannian clustering

Abstract Brain‐Computer Interface (BCI) based on motor imagery (MI) has attracted great interest as a new rehabilitation method for stroke. Riemannian geometry‐based classification algorithms are widely used in MI‐BCI due to their strong robustness and generalization capabilities. However, the clustering performance of current algorithms needs to be improved due to unsuitable clustering criteria for electroencephalography (EEG) characteristics of lower limbs. This study proposed two classification methods based on Riemannian clustering: margin based Riemannian clusters (MBRC) and statistics based Riemannian clusters (SBRC) to address this issue. Our methods divide all samples into subclusters based on the Riemannian distance and innovate clustering criteria. We introduced cluster margin distance and the Riemannian potato algorithm as two clustering criteria to achieve a more robust classification of lower limb MI‐EEG. MBRC and SBRC were tested on an experimental dataset and a public Yi2014 dataset. For the experimental dataset, the average accuracies of MBRC and SBRC were 71.29% and 73.12%, respectively, higher than that of the baseline algorithms. For the Yi2014 dataset, MBRC and SBRC performed better than the comparison algorithms under different training sample numbers, particularly when the number of samples was limited. These findings suggest that the proposed algorithms are more effective for classifying lower limb MI EEGs.