Brain-computer interfaces (BCIs) utilizing motor imagery (MI) have emerged as a promising rehabilitation strategy for individuals recovering from stroke. Classification techniques founded on Riemannian geometry are commonly adopted in MI-BCI systems because of their exceptional robustness and capacity for generalization. Nonetheless, existing algorithms exhibit limitations in clustering performance, primarily due to clustering criteria that are inadequately adapted to the specific characteristics of lower-limb electroencephalography (EEG) signals. To overcome this challenge, the current investigation introduces two novel classification approaches based on Riemannian clustering: margin-based Riemannian clusters (MBRC) and statistics-based Riemannian clusters (SBRC). These methods segment the complete set of samples into subclusters using Riemannian distance measures and introduce original clustering standards. Cluster margin distance and the Riemannian potato algorithm were adopted as the new criteria to enable more reliable classification of lower-limb MI-EEG data. Evaluation of MBRC and SBRC was conducted on both an in-house experimental dataset and the publicly accessible Yi2014 dataset. Results from the experimental dataset revealed average classification accuracies of 71.29% for MBRC and 73.12% for SBRC, exceeding the performance of baseline methods. On the Yi2014 dataset, the proposed algorithms consistently surpassed comparative techniques across different training sample sizes, demonstrating particular advantages under conditions of limited data availability. Overall, these outcomes highlight the enhanced suitability of the developed algorithms for accurate classification of lower-limb MI-EEG signals.
