Characterizing the neural changes following motor rehabilitation in multiple sclerosis from fNIRS signals: a deep learning-enriched analysis

Multiple Sclerosis (MS) is a chronic autoimmune disease of the central nervous system causing progressive motor disability. Understanding cortical changes associated with motor rehabilitation is essential for optimizing treatment, yet the neural mechanisms underlying recovery remain largely unexplored. Functional Near-Infrared Spectroscopy (fNIRS) has been applied to study cortical activation in people with MS (PwMS), though upper-limb task investigations remain limited. This thesis analyzes fNIRS signals recorded during a reach-and-grasp task in 24 PwMS and 8 healthy controls from the PROGR-EX randomized controlled trial. Patients were assigned to three rehabilitation protocols: low-intensity Robot-Assisted Gait Training (LowRAGT), conventional RAGT and Overground Walking Training (OWT). Data were acquired at baseline, post-treatment and three-month follow-up and analyzed through three complementary approaches: traditional statistics, deep learning-based decoding and model interpretability. Statistical results revealed cortical hypoactivation in all patient groups at baseline. The protocols produced distinct trajectories: LowRAGT showed persistent differences from healthy controls with significant ipsilateral cortical reorganization, while conventional RAGT and OWT showed progressive hypoactivation, with OWT exhibiting weaker differences. Among five deep learning architectures tested and compared, a hybrid convolutional-recurrent model achieved the best performance (accuracy up to 0.74), with decoding trends mirroring statistical findings. Interpretability analysis identified the ipsilateral primary motor cortex as the most influential region for discriminating MS from healthy controls. This study provides a preliminary evaluation of the neural effects of three rehabilitation protocols, suggesting that they differ in their capacity to modulate cortical activation in PwMS.