Multimodal Machine Learning to Personalize Transcutaneous Spinal Cord Stimulation for Stroke Rehabilitation.

Introduction: Several studies have identified transcutaneous spinal cord stimulation (tSCS) as a noninvasive neuromodulation technique for improving motor function in individuals with neurological disorders, including stroke. Despite a plethora of preliminary findings, there remains no standardized protocol regarding the optimal tSCS parameters tailored to patients with stroke. The objective of this study was to employ a supervised machine learning (ML) approach to determine the optimal tSCS frequency and intensity parameters for patients with chronic stroke by leveraging data from single-day interventions that systematically varied in frequency and intensity across four stimulation conditions. Methods: Twenty adults with chronic hemiparetic stroke (mean age 53.3 ± 10.8 years; 13 males, 7 females) who were ≥6 months post-stroke was enrolled, excluding individuals with multiple strokes, severe spasticity, or implanted devices. Each participant participated in a baseline session followed by five intervention sessions, during which the frequency and intensity of stimulation were varied randomly. Multimodal sensors, including surface EMG, inertial measurement unit-based kinematics, and spatiotemporal gait parameters, were used to quantify acute changes in gait symmetry, and optimal stimulation frequency and intensity were defined as those yielding the greatest improvement in the combined gait asymmetry metric. A supervised machine-learning classifier was trained using nested leave-one-subject-out cross-validation to predict the optimal stimulation frequency and intensity to maximize differences from the baseline data alone. AUROC values were calculated for the frequency and intensity predictions. Results: The ML models achieved AUROC values of 0.86 [0.75-0.94] for frequency prediction and 0.82 [0.69-0.94] for intensity prediction. The top ten predictive features for each model spanned with spinal motor evoked potentials, wearable sensors, and demographic domains, highlighting multimodal contributions to stimulation optimization. Conclusion: These findings demonstrate that supervised learning can predict individualized tSCS parameters from demographic data and baseline sensor features that yield the greatest improvement in gait symmetry after stroke, representing a promising step toward the data-driven personalization of neuromodulation therapy in neurorehabilitation.