Machine learning–based predictive model for post-stroke cardiorespiratory fitness

BACKGROUND: Reduced cardiopulmonary fitness (CRF) is common in subacute stroke and may limit functional recovery. Timely classification (risk stratification) of CRF status during the subacute phase may help guide rehabilitation when CPET is not readily available. METHODS: 6 relevant features were selected using the Boruta algorithm. Subsequently, nine machine learning (ML) models were developed and evaluated, including logistic regression (LR), elastic net (EN), k-nearest neighbors (KNN), decision tree (DT), extreme gradient boosting (XGB), support vector machine (SVM), random forest (RF), multilayer perceptron (MLP), and Light Gradient Boosting Machine (LightGBM). RESULTS: A total of 114 patients with subacute stroke were included in this study. Among the nine models developed for classifying CRF, RF, SVM, and EN demonstrated the highest discriminative performances, with areas under the receiver operating characteristic (ROC) curve (AUC) of 0.7900, 0.7952, and 0.7905, respectively. Of these models, RF exhibited the greatest clinical applicability. The most important features contributing to classification included age, FMA score, 6MWT distance, sex, lower extremity FMA score, and FAC. CONCLUSION: ML models—particularly RF—most accurately classify concurrent CRF status in subacute stroke using routinely collected variables, supporting clinical triage when CPET is unavailable.