Machine learning score using stress CMR for prediction of all-cause death in patients with suspected or known coronary artery disease a large registry with > 200,000 patient-years of follow-up

Abstract

Traditional prognostic risk assessment in patients undergoing noninvasive imaging is based upon a limited selection of clinical and imaging findings. Machine learning (ML) can consider a greater number and complexity of variables. Therefore, we investigated the feasibility and accuracy of ML to predict 10-year all-cause mortality in patients undergoing stress cardiovascular magnetic resonance (CMR), and compared the performance to existing clinical or CMR metrics. Between 2008 and 2018, a retrospective cohort study with a median follow-up of 6.0 years (IQR: 5.0–8.0) included all consecutive patients referred for stress CMR. The primary outcome was all-cause death based on the National Death Registry. ML involved automated feature selection by random survival forest, model building with a boosted ensemble algorithm, and 10-fold stratified cross-validation. Among the 31,752 consecutive patients (mean age 63.7 ± 12.1 years and 65.7% males), 2679 (8.4%) died at 206,453 patient-years of follow-up. Five clinical and 4 CMR parameters were selected by ML including: age, gender, BMI, history of PAD, renal failure, extent of ischemia and LGE, LVEF and LV end-diastolic volume indexed. Regarding the calibration, C-index on the training sets was 0.730 and 0.735 on the test sets, which attests to the very good robustness of ML models. Machine learning exhibited a higher area-under-curve compared with the FRS, ESC-SCORE and QRISK3 score alone for predicting all-cause mortality (ML score: 0.76 vs. FRS: 0.60, ESC-SCORE: 0.63, QRISK3: 0.62; all P < 0.001) ( Fig. 1 ). Machine learning combining clinical and stress CMR data was found to predict 10-year all-cause mortality significantly better than existing clinical scores alone.