Incremental prognostic value of fully automatic LVEF measured at stress using machine learning

Abstract Background Cardiovascular magnetic resonance (CMR) is the gold standard to measure left ventricular ejection fraction (LVEF), and novel artificial intelligence-based automatic analyses have been proposed for less user interaction and time saving. However, whether automatic LVEF delivers similar information for risk stratification remains unknown. Purpose To assess the prognostic value for all-cause mortality of LVEF measured by stress CMR using a fully automatic machine learning algorithm without human correction. Methods Between 2016 and 2018, all consecutive patients referred for vasodilator stress CMR were included and followed for the occurrence of all-cause death. A fully automatic machine learning algorithm was trained on 3,700 scans and validated on 1,719 unseen CMR studies (MAGNETOM Aera and Skyra, Siemens Healthcare, Erlangen, Germany) to identify end-diastolic and end-systolic phases and segment LV volumes from short-axis cine images at stress. The algorithm combines multiple deep learning networks for detection and segmentation with an active contours approach. Manual and automatic LVEF measured at stress were compared with the paired Wilcoxon test, Pearson correlation and Bland-Altman analysis. Cox regressions were performed to determine the prognostic value of automatic LVEF measured at stress. Results Among 9,883 included patients included to this study, the automatic LVEF was successfully computed in 9,712 (98.3%) patients (66.6% male, mean age 66±12 years). The agreement between manual and automatic LVEF was good (bias = -0.01%, 95% limits of agreement, -6.7% to 6.7%; Pearson’s correlation coefficient r=0.94). A total of 504 (5.2%) deaths were observed during a median (IQR) follow-up period of 4.5 (3.7-5.2) years. Both manual and automatic volumetric assessments showed similar impact on outcome in univariate analyses (manual LVEF per 5%: hazard ratio [HR], 0.80 [95% CI 0.77–0.83]; p<0.001; automatic LVEF per 5%: HR, 0.84 [95% CI, 0.81–0.86]; p<0.001) and multivariable analyses (manual LVEF per 5%: HR, 0.81 [95% CI, 0.78–0.84]; p<0.001; automatic LVEF per 5%: HR, 0.84 [95% CI, 0.82–0.87]; p<0.001). Fully automatic stress LVEF showed an incremental prognostic value to predict all-cause mortality above all traditional risk factors, LVEF measured at rest, the presence of inducible ischemia and LGE (C-statistic improvement: 0.04; NRI=0.221; IDI=0.049; all p<0.001). Conclusions Automatic LVEF is equally predictive of all-cause mortality compared to manual LVEF, and has an incremental prognostic value compared to traditional risk factors, and other stress CMR parameters. Figures: Figure 1: Linear regression plots (A) and Bland-Altman plots (B) comparing stress LVEF measured by experts (LVEFexpert) and computed by fully-automated artificial-intelligence algorithms (LVEFAI). Figure 2: Kaplan-Meier curves for MACE stratified by fully-automated LVEF value. Test comparing the three groups is based on the log-rank test.