Motor-intent decoding from synthetic EEG data using denoising diffusion probabilistic models

Abstract

Decoding motor-intent directly from electroencephalogram (EEG) signals presents significant opportunities for advancing bio-inspired rehabilitation strategies and developing sophisticated human-computer interfaces. Despite the historical dominance of discriminative deep learning decoders, limitations in data availability and the development of effective decoding pipelines remain key obstacles to realizing this potential. This study introduces a novel framework predicated on electromyogram (EMG)-prompted diffusion models for the direct decoding of motor-intent from EMG and EEG signals. We demonstrate that this approach reduces classification error by 12.70% relative to recent discriminative decoders. Furthermore, our results surpassed conventional positive pair augmentation techniques, such as jittering, exhibiting a 3.41% improvement in performance. These findings underscore the transformative potential of generative models for generating synthetic training data and optimizing decoding pipelines in neuro-signal processing. We anticipate that this work will stimulate further investigation into the application of these techniques to improve the efficacy of rehabilitation interventions and facilitate more intuitive human-computer interactions, ultimately contributing to advancements in neuro-assistive device development and personalized rehabilitation strategies.