Stroke rehabilitation requires timely and targeted exercise interventions to restore mobility, strength, and independence. This study develops a machine learning-based system to predict appropriate rehabilitation exercise categories (strength, balance, and mobility) tailored to patient severity levels. Using an open-source dataset of 5,110 stroke patient records, including age, BMI, glucose level, smoking status, paralysis type, and speech ability, three supervised algorithms were evaluated: Random Forest (RF), Logistic Regression (LR), and Multilayer Perceptron (MLP). Accuracy values were reported with 95% Confidence Intervals (CI): RF (94.12%, 95% CI: 93.6–94.6), LR (94.12%, 95% CI: 93.5–94.7), and MLP (94.32%, 95% CI: 93.8–94.9). Despite MLP’s marginally higher accuracy, RF was selected for deployment due to its stability, interpretability, and alignment with expert recommendations. Validation against rehabilitation specialists yielded strong agreement (Cohen’s κ = 0.82), confirming clinical reliability. The RF model was integrated into a web-based application hosted on Heroku. This platform enables patients, particularly those in rural areas with limited access to physiotherapists, to receive personalised exercise guidance. Future work will expand dataset diversity, incorporate hyperparameter optimisation, and evaluate additional metrics such as precision, recall, F1-score, and ROC-AUC to enhance clinical robustness. This system demonstrates the potential of machine learning to support accessible, personalised rehabilitation in resource-constrained settings.
Johan et al. (Wed,) studied this question.