Machine Learning Models to Predict Functional Recovery after Ischemic Stroke: A Nested Cross-Validation Study Using the International Stroke Trial Database

Abstract Predicting functional recovery after ischemic stroke remains challenging due to the heterogeneous nature of cerebrovascular disease. Machine-learning approaches offer potential advantages over traditional statistical methods in capturing complex, non-linear relationships between clinical variables and outcomes. To develop and compare machine-learning models to predict functional recovery in patients with confirmed ischemic stroke using data from the International Stroke Trial (IST). We analyzed 10,016 patients with confirmed ischemic stroke from the IST database. Seven machine-learning algorithms (Logistic Regression, Random Forest, Support Vector Machine [SVM], Stochastic Gradient Descent, XGBoost, Gradient Boosting, and Multilayer Perceptron) were evaluated using nested cross-validation (5 outer folds and 3 inner folds) with probability calibration and Youden index threshold optimization. Six resampling strategies were compared to address class imbalance (28% of recovery rate). Model interpretability was assessed using Shapley additive explanations (SHAP) values. Logistic regression achieved the highest area under the receiver operating characteristic curve (AUC-ROC: 0.647; 95%CI: 0.636–0.658) without resampling, demonstrating comparable performance to that of more complex algorithms. The best sensitivity (0.677; 95%CI: 0.620–0.733) was achieved by XGBoost with Synthetic Minority Over-sampling Technique (SMOTE)-Tomek resampling, while SVM with Tomek Links achieved the highest specificity (0.740; 95%CI: 0.700–0.781). Resampling strategies improved sensitivity at the cost of specificity, without significant improvement in discriminative ability. Key predictive features included leg/foot motor deficit, stroke subtype, hemianopia, age, and dysphasia. Machine-learning models demonstrated modest discriminative ability to predict functional recovery after ischemic stroke. Logistic regression provided the best balance between performance and interpretability. Additional variables, such as imaging biomarkers, may be necessary to improve prognostic accuracy.