Machine learning–based prediction of activities of daily living in stroke patients using lower limb rehabilitation robot training data

Objectives: This study aimed to predict activities of daily living (ADL) outcomes in stroke patients undergoing lower-limb robotic rehabilitation using multiple machine learning algorithms, to support personalized and precise rehabilitation strategies. Methods: A retrospective analysis was conducted on the baseline data of 618 stroke patients who received conventional lower-limb rehabilitation robot training. ADL was assessed using the Barthel Index (BI), with admission BI scores ≤60 labeled “A” and >60 labeled “a”; BI improvement ≤15 labeled “B” and >15 labeled “b,” resulting in 4 recovery groups. Features were selected via Lasso regression, and data were split (70:30) into training and test sets. Seven classification algorithms—Multinomial Logistic Regression, Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Naive Bayes, K-Nearest Neighbors (KNN), and Multilayer Perceptron (MLP)—were developed and evaluated. Predictor importance was ranked across baseline ADL levels. Model performance, clinical utility, and robustness were assessed using decision curve analysis (DCA) and reclassification metrics. Results: Functional Ambulation Category (FAC) and fibrinogen (FIB) consistently ranked highest in predictor importance. The Multilayer Perceptron (MLP) model achieved the best overall performance (accuracy=0.6902; precision=0.6769, and F1=0.7629). Multiclass receiver operating characteristic (ROC) analysis showed the strongest discriminative power for the “ab” group (XGBoost AUC=0.972; MLP AUC=0.953; and RF AUC=0.944). DCA suggested “AB vs. Rest” is optimal for high-risk screening, while “Ab vs. Rest” supports intervention prioritization. Conclusions: This study establishes a quantitative framework for assessing ADL recovery in stroke patients receiving lower-limb robotic rehabilitation, informs early clinical intervention strategies, and offers insights into predictive model optimization.