Improving Motor Imagery Decoding in Elderly by Combining Cross-Population Model With Tactile-Induced Sensorimotor Rhythm

In motor imagery-based Brain-Computer Interfaces (MI-based BCIs), research has predominantly focused on developing algorithms and paradigms optimized for young adults. However, recent studies have revealed a significant decline in algorithm performance when applied to elderly subjects, posing a major challenge to the practical deployment of MI-BCIs in this demographic. To address this limitation, we propose a novel cross-population transfer learning-based decoding scheme. Specifically, we leverage sensory and motor multi-modal hybrid data to fine-tune a pre-trained model, originally trained on MI data from young adults. To further mitigate inter-subject variability in EEG signals, we introduce a cross-task Euclidean alignment strategy, where MI data is aligned using covariance computed from sensory data, enabling MI decoding for elderly subjects without requiring any MI data from the target population. Our proposed approach has demonstrated promising results, achieving a 14% improvement in decoding performance compared to within-subject decoding and a 6.64% enhancement over cross-subject decoding for elderly participants. Additionally, our method reduces the BCI-system-deficiency rate by 47.37% compared to within-subject decoding methods, and by 15.79% compared to the strongest cross-subject baseline, significantly enhancing system robustness and reliability. By eliminating the need for subject-specific MI data, our approach provides a plug-and-play solution for clinical applications, particularly in stroke rehabilitation and elderly neurorehabilitation, paving the way for more inclusive, accessible, and adaptable BCI technologies.