300 Section - Predicting Motor and Activities of Daily Living Outcomes in Stroke Rehabilitation Using a Multi-Modal Longitudinal Dataset and Machine Learning

Objectives : Accurate prediction of recovery after stroke is crucial for tailoring rehabilitation but remains a significant clinical challenge. This study leveraged a multi-dimensional, longitudinal dataset to develop and validate machine learning models for predicting upper-extremity motor function and Activities of Daily Living (ADL) outcomes in stroke survivors.

Design: We prospectively enrolled 120 stroke patients undergoing a standardized upper-limb rehabilitation program. Data were collected at baseline and post-treatment, encompassing patient history, comprehensive clinical assessments (e.g., Fugl-Meyer Assessment [FMA], Modified Barthel Index), and neurophysiological measures (Motor Evoked Potentials, MEP). We employed three machine learning algorithms (Support Vector Machine, Random Forest, XGBoost) to predict post-treatment FMA and ADL scores using various feature combinations. Model performance was assessed using the R² coefficient, and feature importance was determined via SHapley Additive exPlanations (SHAP).

Results: Correlation analysis confirmed high convergent validity (r > 0.7) among established motor function scales. Machine learning models predicted FMA scores with high accuracy; the Random Forest model achieved a robust R² of 0.82 using a combination of patient history and baseline MEP data. In contrast, predicting ADL outcomes was more complex, with the optimal model (SVM using basic patient information) yielding an R² of 0.68. SHAP analysis revealed that MEP features, while being powerful predictors, were prone to overfitting. Other key predictors included Body Mass Index (BMI), history of hypertension, and the specific sequence of therapeutic interventions. The results also quantitatively confirmed that patients' ADL performance was more strongly correlated with the function of their unaffected hand, highlighting the significant role of compensatory strategies.

Conclusion: This study demonstrates that machine learning models incorporating multi-modal data can effectively forecast post-stroke motor recovery. Upper-extremity motor function, closely linked to corticospinal tract integrity, is more predictable than global ADL. These findings provide a data-driven basis for personalizing rehabilitation strategies and advancing precision in stroke recovery