Risk stratification in patients with known coronary artery disease (CAD) remains a clinical challenge, especially in asymptomatic individuals. While stress cardiac magnetic resonance imaging (MRI) has strong prognostic value, current models do not fully exploit the richness of available clinical and imaging data. Machine learning (ML) offers an opportunity to optimize prediction by capturing complex patterns in high-dimensional datasets.

To assess the performance of a supervised ML model combining clinical and stress cardiac MRI data for predicting 10-year major adverse cardiovascular events (MACE) in asymptomatic patients with obstructive CAD, compared to logistic regression models.

A total of 966 asymptomatic patients with obstructive CAD who underwent vasodilator stress cardiac MRI between 2009 and 2011 in two centres were retrospectively included. The first centre (n = 742) provided a derivation cohort (n = 603) and an internal validation cohort (n = 139), while the second centre (n = 224) served as an external validation cohort. Feature selection was performed using LASSO, XGBoost, Random Forest (RF), and Boruta. A final RF model was trained using five selected variables and compared to a generalized logistic regression model (GLM) using AUROC and PRAUC metrics.

Five key variables were selected: number of ischemic segments, number of late gadolinium enhancement (LGE) segments, left ventricular ejection fraction (LVEF), left ventricular end-diastolic diameter indexed, and age ( Fig. 1 ). Over the 10-year follow-up period, MACE occurred in 28% of patients in the derivation cohort, 30% in the internal validation cohort, and 24% in the external test cohort. The RF model demonstrated the best predictive performance in the derivation cohort (AUROC: 0.99, PRAUC: 0.98). Consistently, the AUROC was 0.98 versus 0.86 for the GLM, and PRAUC was 0.97 versus 0.78 (all p < 0.001) in the internal validation cohort. In the external validation cohort, the RF model achieved an AUROC of 0.92 versus 0.74 for the GLM, and a PRAUC of 0.84 versus 0.58 (all p < 0.001). SHAP analysis confirmed the interpretability of the model and the individual contribution of each variable ( Fig. 2 ).

A ML model combining stress cardiac MRI and clinical data significantly outperformed traditional methods in predicting MACE in asymptomatic patients with obstructive CAD.