一种用于提升运动想象脑机接口性能的协同自适应少样本校准方法

Motor Imagery Brain-computer Interfaces connect the human brain with external devices by imagining muscular activity, serving as one of the promising solutions for human-computer interaction. It has been successfully applied by paralyzed individuals to control assistive devices, prosthetics and exoskeletons, and several dedicated applications. However, about half of potential users exhibit MI illiteracy, which hinders its adoption in real-world settings. Calibration has been demonstrated as an effective approach to enhancing MI capabilities. Nevertheless, existing calibration methods face two critical challenges: 1) the lack of systematic protocols to guide subjects in refining their MI skills based on the outcome of the intended actions, and 2) the use of fixed, cross-subject classifiers that do not address inter-subject MI variability. This study proposes the Co-adaptive Few-shot calibration (CFC) approach. It particularly addresses the above two challenges: 1) it introduces MI pattern selection to provide a subject-specific assessment protocol, and 2) it incorporates meta-learning in the calibration stage to address inter-subject and inter-pattern variability and reduce calibration time. We carried out an experiment on 30 subjects to verify the effectiveness of CFC. The results show that after a short calibration, the subject's MI capabilities were significantly improved, with the average binary motor imagery accuracy improved from 69.50% to 77.47% (+ 7.97%) on the experiment group with 10 naive subjects. The CFC approach can help subjects find the optimal MI pattern to enhance their MI ability in a short time. This method broadens the practical usability of MI-BCI systems, making them accessible to a wider population of users.