SHAP-Assisted Machine Learning Model to Predict Stroke Location in Atrial Fibrillation-Related Ischemia

The anatomical distribution of ischemic stroke in patients with atrial fibrillation (AF) is highly heterogeneous, yet its biological determinants remain poorly defined. This study aimed to identify clinical and molecular predictors of stroke topography and to develop explainable machine learning models to forecast lesion location. We retrospectively analyzed 500 patients with AF-related ischemic stroke, collecting clinical data and blood-based biomarkers including C-reactive protein (CRP), interleukin-6 (IL-6), B-type natriuretic peptide (BNP), D-dimer, and fibrinogen. Patients were stratified by stroke territory (anterior vs. posterior circulation) and hemispheric side (left vs. right). Two XGBoost models were built to predict each spatial pattern. Model performance was assessed using ROC analysis, and SHapley Additive exPlanations (SHAP) were applied for feature interpretation. Posterior circulation infarcts were associated with higher CRP and IL-6 levels (p < 0.0001), while BNP was elevated in left-sided strokes, and IL-6 in right-sided strokes. The circulation model achieved an AUC of 0.71, and the hemispheric model an AUC of 0.76. SHAP highlighted CRP and IL-6 as top predictors for posterior infarcts, and BNP for left-sided infarcts. These findings suggest that inflammatory and cardiac biomarkers carry spatially specific predictive value in AF-related stroke and support the use of interpretable models for personalized risk stratification.