Key result
AI-derived cardiac MRI EF outperforms clinician EF in predicting all-cause mortality across CAD and NICM.
Why the study?
The authors hypothesized that deep-learning AI segmentation of EF and other LV volume metrics from cardiac MR outperforms expert clinician segmentation for predicting all-cause mortality.
Does deep-learning AI segmentation of cardiac MRI improve the prediction of all-cause mortality compared to expert clinician segmentation?
Does deep-learning AI segmentation of cardiac MRI improve the prediction of all-cause mortality compared to expert clinician segmentation?
Deep-learning AI segmentation of cardiac MRI outperforms expert clinician segmentation in predicting all-cause mortality, supporting its use as a fast, low-cost alternative for risk stratification.
Background Ejection fraction (EF), a key measure of cardiac function, is used in healthcare for clinical decision-making and risk stratification. We hypothesized that a deep-learning AI segmentation of EF and other LV volume metrics obtained from cardiac MR outperforms expert clinician segmentation for the prediction of all-cause mortality. Methods We studied a retrospective cohort of 27,382 consecutive patients who underwent CMR in the UK (n=26,106 between 2014-2019) & USA (n=1,276 between 2010-2016). Mortality data was obtained from national and local databases. Median follow-up was 5.6 years (IQR: 4.4-6.9 years) with 3,458 (13%) deaths. We performed AI segmentation on all CMR studies and fitted Cox Proportional-Hazards models. We also studied a subgroup of 8,742 patients who had either coronary artery disease (CAD, n=4,933), non-ischaemic cardiomyopathy (NICM, n=2,198) and hypertrophic cardiomyopathy (HCM, n=1,611). Results Compared to clinician segmentation, AI EF had higher prediction for all-cause mortality (p<0.001). Differences remained after adjusting for age and sex. AI left ventricular mass index (LVMi) and left ventricular end diastolic volume indexed (LVEDVi) also had higher prediction (both p<0.001). Prognostic models that included EF, LVEDVi, and LVMi had higher prediction than EF alone, and in these models, AI was also superior to clinicians. AI EF was superior to clinicians in CAD (p<0.001) and NICM (p=0.027). Differences remained when adjusting for the presence of viability and ischaemia in CAD and focal fibrosis in NICM. LV volumetric parameters were poorly predictive of mortality in HCM (all C-Indices <0.6) and in these clinician EF had slightly higher prediction (p=0.002). Conclusion AI-derived EF shows superior prediction of mortality compared to clinicians in all-comer patients and in patients with CAD and NICM. AI EF is deployed at high speed and low-cost. Findings support the superseding of clinician EF with AI.
No takes yet. Share an insight, caveat, or question.
Joy et al. (2025) studied this question. AI-derived ejection fraction outperformed clinician EF in predicting all-cause mortality in 27,382 cardiac MRI patients, with p<0.001 overall and in CAD and NICM subgroups.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: