Current MI-BCI methods face several limitations, according to the authors: low generalizability caused by cross-patient variability, the absence of an effective functional assessment step, limited availability of patient data, and a lack of suitable data augmentation approaches. The proposed framework introduces a diffusion model tailored to the spatio-temporal characteristics of EEG, incorporating a decoupled neural architecture with rotary spatial encoding and autoregressive temporal fusion. Two augmentation strategies adapted to the characteristics of stroke EEG were designed to mitigate data scarcity.
Extensive experiments across multiple MI-BCI tasks demonstrate superior performance and generalizability, the authors say, supporting the potential of the method for deployment in personalized stroke rehabilitation. The abstract reports no specific accuracy figures and no prospective clinical validation.