Abstract Number: Esoc2026lb112 Development and Validation of a Machine Learning Model for Predicting One-Year Stroke Recurrence With Etiologic and Severity Stratification

Abstract Background and aims Stroke recurrence is clinically heterogeneous, yet prediction models often use binary outcomes that overlook etiologic mechanisms and severity. We aim to develop subtype-specific models for one-year recurrence and a severity-aware framework to predict fatal recurrence. Methods We analyzed data from the CASE-II registry, a prospective, multicenter stroke registry in China. For subtype-specific modeling, we implemented a two-step XGBoost approach with baseline training followed by TOAST subtype fine-tuning. For recurrence severity, we classified five outcomes based on post-recurrence mRS (no, mild, moderate, severe, fatal recurrence mRS 6) by training a multiclass XGBoost model on the same baseline predictors. Key predictors for fatal recurrence were identified via SHAP analysis and integrated into a Fine-Gray competing risk model, which was simplified into a 10-item clinical risk score. Results Among 20,756 patients, 2,919 (14.06%) experienced stroke recurrence within one year. Subtype-specific models significantly outperformed the baseline model (AUC=0.53), achieving AUCs of 0.70(CE), 0.65(LAA), 0.69(SAA), 0.89(SOE), and 0.68(SUE). For recurrence severity prediction, the multiclass model demonstrated good discrimination for fatal recurrence (AUC=0.79). The final 10-item score, incorporating age, NIHSS, and other biomarkers, stratified patients into low-(34.2% of cohort, event rate 0.49%, HR 1.0), intermediate-(51.9%, event rate 2.64%, HR 5.4), and high-(13.9%, event rate 6.96%, HR 14.6) risk groups for fatal recurrence. Conclusions Etiologic subtype stratification improves stroke recurrence prediction. Furthermore, we created a severity-aware 10-item score identifying a high-acuity population with 15-fold increased risk for fatal recurrence, facilitating triage for intensified secondary prevention following ischemic stroke. Conflict of interest