AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
XGBoost Outperforms Traditional Risk Scores for Five-Year MACE Prediction in a Moroccan Multicenter Cohort: An Artificial Intelligence Based Approach with SHAP Explainability
By Nguadi J, elboussaadani B, fagouri J et al.
BACKGROUND: Traditional cardiovascular risk scores (SCORE2, PREVENT, Framingham) systematically underestimate risk in North African populations due to the high prevalence of poorly controlled diabetes, chronic kidney disease, and hypertension. Cardiovascular disease remains the leading cause of mortality in Morocco, accounting for 38% of all deaths. There is an urgent need for population-specific predictive tools that integrate cardiorenal variables and capture non-linear interactions. Research QUESTION / HYPOTHESIS: We hypothesized that an XGBoost model trained on a real-world Moroccan multicenter cohort would significantly outperform traditional risk scores in predicting 5-year major adverse cardiovascular events (MACE), with superior AUC-ROC and Recall, and provide clinically interpretable predictions through SHAP explainability. MethodS: Multicenter retrospective study across 7 Moroccan cardiology centers (May 2019 - January 2026). We included 623 adults ≥40 years with median follow-up of 4.8 years. Variables included clinical, biological, echocardiographic data, plus engineered features (ageхdiabetes, eGFRхLVEF). XGBoost model with hybrid ROS-SMOTE, GridSearchCV optimization, temporal split + 10-fold CV + bootstrap (n=2000). Compared to SCORE2, PREVENT, and Framingham. SHAP for explainability. Primary metrics: AUC-ROC, Recall, NRI, Brier score. ResultS: Mean age 59.1 ± 11.5 years; diabetes 44%; hypertension 68%; eGFR <60 mL/min 31%; MACE rate 22%. XGBoost achieved AUC-ROC 0.88 (95% CI 0.85-0.91), Recall 0.84, Precision 0.61, Brier score 0.08, and NRI +36% versus SCORE2. SCORE2 and PREVENT showed AUC-ROC of 0.71 and 0.67, respectively (Figure 1). SHAP identified age, LVEF, eGFR, and HbA1c as top predictors with non-linear effects in renal impairment (Figure 2a, 2b). Calibration and Decision Curve Analysis showed superior net benefit at 5-20% thresholds. Conclusions: XGBoost significantly outperforms SCORE2 and PREVENT for 5-year MACE prediction in this real-world Moroccan cohort (AUC-ROC 0.88 vs 0.71; Recall 0.84 vs 0.62; NRI +36%; p<0.001). SHAP explainability supports clinical adoption. After external validation, this model could improve cardiovascular prevention in Morocco and similar middle-income settings.