Development and validation of interpretable machine learning models for predicting stroke in NVAF patients with CHA

Background: Non-valvular atrial fibrillation (NVAF) patients with CHA₂DS₂-VA ≤1 face uncertainty in stroke risk assessment, particularly in Asian populations. Machine-learning (ML) models offer improved accuracy for individualized risk prediction. Methods: This single-center, retrospective study at Zhongshan Hospital, Xiamen University (January 2022-January 2025) included 82 NVAF-related stroke cases and 164 matched non-stroke controls with CHA₂DS₂-VA ≤1. Data encompassed demographics, comorbidities, laboratory markers, and echocardiographic parameters. Following stratified train-test split (80%:20%), feature selection used univariable logistic regression and least absolute shrinkage and selection operator (LASSO) regression, retaining the intersection of variables selected by both methods. Three nested predictor sets were defined: Model A (clinical and routine laboratory variables), Model B (Model A plus Cardiac biomarkers), and Model C (Model B plus echocardiographic parameters). ML algorithms [logistic regression (LR), random forest (RF), XGBoost (XGB)] underwent nested cross-validation for hyperparameter tuning. Performance was evaluated by area under the receiver operating characteristic curve (AUC), sensitivity, specificity, precision, and F1 score. Shapley additive explanations (SHAP) were applied to the best-performing ML algorithm in the test set to evaluate the contributions of individual features. Results: Stroke cases were older, with higher E/e' ratio, and elevated N-terminal pro-B-type natriuretic peptide (NT-proBNP), C-reactive protein (CRP), and white blood cell count (all Conclusions: ML models incorporating cardiac biomarkers and echocardiographic parameters improve stroke risk stratification in low- to moderate-risk NVAF patients, supporting personalized anticoagulation strategies.