Toward personalized neurorehabilitation: Machine learning approach in predicting treatment response in stroke patients

Stroke, also knownascerebrovascularaccident(CVA),isaleadingcauseoflong-termdisability worldwide. According to the World Health Organization, approximately 15 million people experience a stroke each year, with nearly five million living with permanent impairments. Upper-limb motor deficits are particularly debilitating, affecting up to 85% of survivors in the early post-stroke phase, with around 60% experiencing persistent limitations beyond six months, substantially reducing independence, quality of life, and imposing significant social and economic burdens. Despite advances in acute stroke care and rehabilitation, patient responses remain highly variable due to factors such as lesion characteristics, comorbidities, and individual neuroplasticity, leaving clinicians with considerable uncertainty in tailoring therapy. This thesis addresses this gap by showing that integrating multimodal physiological data with machine learning enables accurate, individualized prediction of motor recovery. Using biomarkers from surface electromyography, kinematics, and clinical assessments, these frameworks aim to guide therapy selection, identify patients most likely to benefit, and maximize functional recovery while minimizing resource use and patient burden. The thesis begins by demonstrating the feasibility of predicting upper limb motor recovery in post-stroke patients undergoing robot-assisted rehabilitation with the ALEx exoskeleton. In a pilot study of nine subacute stroke survivors, multimodal features were extracted from surface EMG recordings of fifteen upper-limb muscles and from kinematic data acquired during three-dimensional reaching tasks. Machine-learningclassifiers wereusedtodistinguish recoverers from non-recoverers based on changes in Fugl-Meyer Assessment scores after a four-week robotic intervention. The median contraction amplitude of the brachioradialis muscle emerged as the most predictive single feature, achieving subject-level classification accuracies above 95% and providing clinically meaningful prognostic information. These findings establish the methodological foundation for the subsequent studies presented in this thesis. Building upon these results, the thesis expands the predictive framework to incorporate both regression modeling and anomaly detection in eleven subacute stroke patients, evaluating model performance with and without baseline Fugl-Meyer Assessment scores to assess the feasibility of prediction using only objective physiological data. A novel two-step machine-learning pipeline was devised: regression models first estimated post-intervention functional scores, while a fully connected autoencoder subsequently identified patients with limited or no recovery through anomaly detection. This dual approach addressed the inherent class imbalance between recoverers and non-recoverers, providing both precise quantitative predictions and a mechanism to flag unreliable estimates. A key innovation was the use of the Gesture Similarity Index, a compact multimodal metric integrating muscle synergy patterns and movement smoothness to quantify similarity to normative movement. The multimodal models predictions achieved median root mean square errors below the minimal clinically important difference for most patients, and models excluding baseline scores showed comparable performance. The anomaly detection component achieved an F1-score above 0.98 in identifying non-responders, supporting robust identification of patients who may require alternative therapeutic strategies. Toevaluate generalizability beyond controlled robotic environments, the thesis extends the predictive framework to wearable technology and physiotherapist-guided hand rehabilitation. Thirteen chronic stroke patients participated in a three-week task-oriented hand therapy program while wearing the REMO® device, which simultaneously records electromyographic signals from eight sEMG channels around the forearm and inertial measurement data. Four families of machine learning models were trained to predict post-intervention Fugl-Meyer Assessment hand subscores, systematically assessing the contribution of baseline clinical scores and the Gesture Similarity Index. Models excluding baseline assessments achieved median prediction errors comparable to those including baseline scores, with over 75% of participants exhibiting errors below clinically meaningful thresholds. The Gesture Similarity Index consistently emerged as a dominant predictive feature, highlighting its potential as a data-driven surrogate for conventional clinical assessments. These results demonstrate that accurate, individualized predictions can be achieved using lightweight wearable devices in real-world rehabilitation settings, substantially broadening the clinical applicability and scalability of predictive modeling approaches. Complementing the focus on predictive modeling, this final chapter investigates the acute physiological effects of transcutaneous spinal cord stimulation (tSCS) as a potential neuromodulatory intervention for stroke rehabilitation. Thirteen chronic stroke patients completed six experimental sessions with varying stimulation frequencies and intensities. Spinally evoked potentials recorded from ten lower-limb muscles revealed that single-session tSCS selectively modulated temporal conduction properties and subject-specific tolerance, with the most pronounced effects observed at 50 Hz and 30% of resting motor threshold. Latency reductions were greatest in the tibialis anterior and hamstrings, indicating enhanced responsiveness of spinal motor neuron pools. In contrast, amplitude-based measures and resting motor thresholds remained largely unchanged, suggesting that acute tSCS preferentially improves conduction efficiency rather than motoneuron recruitment magnitude. These results provide quantitative guidance for selecting stimulation parameters in future therapeutic applications. In a side project presented in the supplementary materials, the thesis further extended the scope of data-driven assessment by investigating whether machine learning methodologies developed for post-stroke rehabilitation could be applied to other neurological populations. Acoustic features extracted from speech tasks were used to classify speech and swallowing impairment in 102 patients with amyotrophic lateral sclerosis. K-nearest neighbors classifiers achieved test accuracies of 65% for speech impairment and 70% for swallowing impairment, substantially exceeding chance-level performance. Local jitter during syllable repetition emerged as the most predictive feature for speech impairment, while harmonics-to-noise ratio was most informative for swallowing dysfunction. These findings demonstrate the versatility and generalizability of multimodal machine learning frameworks across clinical populations and assessment contexts. In conclusion, this thesis demonstrates that the convergence of multimodal physiological data acquisition, machine learning, and advanced rehabilitation technologies provides a promising pathway toward precision, individualized neurorehabilitation. By enabling clinicians to anticipate recovery potential, optimize therapy selection, and incorporate evidence-based neuromodulatory strategies, this work transforms rehabilitation from a largely empirical process into a scientifically guided, patient-centered practice. The data-driven frameworks developed hold substantial promise for enhancing rehabilitation outcomes, reducing healthcare costs, and improving quality of life for stroke survivors worldwide.