Machine-learning score using stress CMR and CCTA for prediction of cardiovascular events in patients with obstructive CAD

Abstract

In patients with suspected or known CAD, traditional prognostic risk assessment is based upon a limited selection of clinical and imaging findings. Machine learning (ML) methods can take into account a greater number and complexity of variables. To investigate the accuracy of ML-score using simultaneously stress CMR, coronary CT angiography (CCTA), and clinical data to predict the occurrence of CV events in patients with suspected or known CAD. Between 2008 and 2020, consecutive symptomatic patients without known CAD referred for CCTA were screened in ICPS (Massy). Patients with obstructive CAD (at least one ≥ 50% stenosis on CCTA) were further referred for stress CMR and followed for the occurrence of major adverse cardiovascular events (MACE), defined as CV death or nonfatal myocardial infarction. Twenty-three clinical, 11 stress CMR and 11 CCTA parameters were evaluated. ML involved automated feature selection and model building by random survival forest. The external validation cohort was Lariboisiere Hospital (N = 274 patients). Of 2038 consecutive patients (47% men; mean age 69 ± 12 years), 281 (13.8%) patients experienced a MACE after a median follow-up of 6.7 years (interquartile range: 5.9–9.1). Our ML score exhibited a higher area-under-the-curve compared with stress CMR data alone, CCTA data alone, and traditional Cox model for prediction of 10-year MACE (ML: 0.88 vs. CMR data alone: 0.79, CCTA data alone: 0.72; traditional Cox model: 0.81, all P < 0.001). The ML score assessed in the derivation cohort (AUC: 0.88, F1-score 0.80) exhibited also a good area-under-the-curve in the external cohort for prediction of 10-year MACE (AUC: 0.86, F1-score 0.80). The ML score including clinical, stress CMR and CCTA data exhibited a higher prognostic value to predict 10-year MACE compared with all traditional clinical data, CMR data or CCTA data alone (Fig. 1).