Uncovering Low-Dimensional Manifolds of Neural Dynamics for Motor-Imagery Based Stroke Rehabilitation: An EEG-Based Brain–Computer Interface Study

Abstract

Stroke rehabilitation aims to repair neural circuits and dynamics through the remapping of neuronal functions. However, there is currently a gap in understanding the alteration of neural population dynamics-the fundamental computational unit driving functions-under clinical settings. In this study, we introduced a novel method to identify stable low-dimensional structures of neural population dynamics in stroke patients during motor tasks. Using whole-brain EEG recordings from chronic stroke patients performing motor imagery (MI) tasks before and after brain-computer interface (BCI) training, as well as a public EEG dataset of acute stroke patients performing MI tasks, we projected EEG signals from sensor space to voxel space via source localization (eLORETA), simulating neural population activity in regions of interest. By applying dimensionality reduction, we successfully obtained low-dimensional neural manifolds to represent neural population dynamics. Our analysis revealed three key findings: (1) For right-handed patients, task-related low-dimensional dynamics in the related brain regions remain stable across subjects, with their features holding potential as biomarkers for stroke rehabilitation; (2) BCI training promotes global and sustained restoration of neural population dynamics; (3) EEG theta-band oscillations show strong correlation with these dynamics, highlighting their macroscopic nature. This study proposes a new, simple, and powerful tool for comprehension and validation of stroke rehabilitation mechanisms confirming the effectiveness of BCI training in restoring neural dynamics.