Machine-Learning Score Using Stress CMR for Death Prediction in Patients With Suspected or Known CAD

Abstract

Background: In patients with suspected or known coronary artery disease, traditional prognostic risk assessment is based on a limited selection of clinical and imaging findings. Machine learning (ML) methods can take into account a greater number and complexity of variables.

Objectives: This study sought to investigate the feasibility and accuracy of ML using stress cardiac magnetic resonance (CMR) and clinical data to predict 10-year all-cause mortality in patients with suspected or known coronary artery disease, and compared its performance with existing clinical or CMR scores.

Methods: Between 2008 and 2018, a retrospective cohort study with a median follow-up of 6.0 (IQR: 5.0-8.0) years included all consecutive patients referred for stress CMR. Twenty-three clinical and 11 stress CMR parameters were evaluated. ML involved automated feature selection by random survival forest, model building with a multiple fractional polynomial algorithm, and 5 repetitions of 10-fold stratified cross-validation. The primary outcome was all-cause death based on the electronic National Death Registry. The external validation cohort of the ML score was performed in another center.

Results: Of 31,752 consecutive patients (mean age: 63.7 ± 12.1 years, and 65.7% male), 2,679 (8.4%) died with 206,453 patient-years of follow-up. The ML score (ranging from 0 to 10 points) exhibited a higher area under the curve compared with Clinical and Stress Cardiac Magnetic Resonance score, European Systematic Coronary Risk Estimation score, QRISK3 score, Framingham Risk Score, and stress CMR data alone for prediction of 10-year all-cause mortality (ML score: 0.76 vs Clinical and Stress Cardiac Magnetic Resonance score: 0.68, European Systematic Coronary Risk Estimation score: 0.66, QRISK3 score: 0.64, Framingham Risk Score: 0.63, extent of inducible ischemia: 0.66, extent of late gadolinium enhancement: 0.65; all P < 0.001). The ML score also exhibited a good area under the curve in the external cohort (0.75).

Conclusions: The ML score including clinical and stress CMR data exhibited a higher prognostic value to predict 10-year death compared with all traditional clinical or CMR scores.

Keywords: all-cause mortality; cardiac magnetic resonance; ischemia; machine learning; stress testing.