Fully automated assessment of global longitudinal strain by machine learning predicts death in patients undergoing stress CMR

Abstract Background Several studies described the independent prognostic value of left ventricular global longitudinal strain (GLS) using cardiovascular magnetic resonance (CMR) to predict death and cardiovascular events. While routine GLS evaluation is a time consuming and partially operator-dependent process, the potential interest of GLS using a fully automated method has not yet been established. Purpose To assess the feasibility and prognostic value of GLS to predict all-cause death using a fully automated machine-learning algorithm without human correction in consecutive patients referred for stress CMR. Methods Between 2016 and 2018, all consecutive patients referred for stress CMR were included and followed for the occurrence of all-cause death. A fully automated machine learning algorithm was trained and validated on unseen CMR studies (MAGNETOM Aera and Skyra) to assess the GLS from long-axis cine images acquired at rest. The algorithm combines multiple deep learning networks for detection and segmentation with an active contours approach. Cox regressions were performed to determine the prognostic value of GLS. Results In a retrospective study of 9,883 consecutive patients who underwent stress CMR between 2016 and 2018, automated GLS was successfully computed in 9,638 (97.5%) patients (67% male, mean age 66±12 years). A total of 510 (5.3%) deaths were observed during a median follow-up period of 4.5 years (interquartile range: 3.7-5.2). GLS, the presence of inducible ischemia and late gadolinium enhancement (LGE) were significantly associated with the occurrence of death (hazard ratio, HR: 1.22 [95% CI, 1.17-1.26]; HR: 2.23 [95% CI, 1.61-3.10]; and HR: 2.04 [95% CI, 1.41-2.95], respectively, all p<0.001). After adjustment for traditional risk factors, inducible ischemia and LGE, GLS was an independent predictor of a higher incidence of death (adjusted HR: 1.14 [95% CI, 1.08-1.20]). Automated GLS showed an incremental prognostic value to predict death compared to traditional risk factors, inducible ischemia and LGE (C-statistic improvement: 0.05; NRI=0.146; IDI=0.244; all p<0.001). Conclusions Fully automated GLS measured at rest has an incremental prognostic value to predict all-cause death above traditional risk factors, and other stress CMR parameters. This finding suggests that unsupervised GLS measurements are accurate and allow improved risk stratification.