New AI Framework PA-TCNet Improves Stroke Rehabilitation Brain-Computer Interface Decoding

A novel AI framework called PA-TCNet has been developed to address challenges in cross-subject electroencephalography decoding for stroke patients undergoing motor imagery brain-computer interface rehabilitation. The system specifically tackles lesion-related abnormal temporal dynamics and significant inter-patient variability that hinder generalization in existing adaptation methods. PA-TCNet incorporates two coordinated components: the Pathology-aware Rhythmic State Mamba module decomposes EEG spatiotemporal features into slow rhythmic context and fast transient perturbations, while a physiology-guided target refinement mechanism helps stabilize unreliable target-domain pseudo-labels. This approach aims to more effectively capture pathological slow-wave activity that often misleads current adaptation techniques. The framework represents an advancement in motor rehabilitation technology for stroke survivors by improving the decoding of motor imagery EEG signals across different patients. Research detailing PA-TCNet was published on arXiv under identifier 2604.16554v1.