Predicting 10-year Major Adverse Cardiac Events Using Multi-Source Modalities with XGBoost: Establishing a Baseline for Multimodal Fusion in Cardiac Risk Assessment
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Accurate prediction of major adverse cardiac events (MACE) is critical for long-term cardiovascular risk management. Traditional risk scores offer only moderate performance. Leveraging multi-source data may improve individualized risk stratification.
Methods
In this retrospective study of patients who underwent non-contrast cardiac- gated CT between 2010 and 2023 across Emory-affiliated hospitals, XGBoost models were trained on structured tabular data using sequential feature integration to predict 10-year MACE. Features included coronary artery calcium (CAC), other imaging- derived metrics, clinical risk scores, electrocardiogram parameters, and laboratory biomarkers. Performance was mainly assessed using AUC-ROC and AUC-PRC. A 5- fold cross-validation strategy was employed, repeated across 10 randomized seeds. Statistical significance was evaluated using two-sided t-tests with 95% confidence.
Results
This retrospective study included 25,514 adult patients (mean age 57 ± 10 years; 57% men), of whom 2.93% experienced MACE within 10 years. The final model incorporating all features, achieved the highest performance with an AUC-ROC of 0.883 ± 0.012, a 30.8% improvement over CAC (0.675 ± 0.015), 28.9%-32.2% over clinical risk scores, with p <0.01 for all. AUC-PRC was 0.289 ± 0.028 compared to 0.056-0.104 for clinical risk scores and 0.067 for CAC. SHAP analysis identified creatinine, hemoglobin A1c, body mass index, glomerular filtration rate, and CAC volume as the most influential features.
Conclusion
Sequential integration of structured clinical and imaging-derived data significantly improves MACE prediction. This model establishes a robust and interpretable benchmark for future research in multimodal fusion and cardiovascular risk stratification.
