Abstract Body (Do not enter title and authors here): Background: Coronary artery calcium (CAC) computed tomography (CT) is widely used for atherosclerotic cardiovascular risk assessment but is not routinely utilized for predicting incident heart failure (HF). Deep learning now enables automated extraction of cardiac morphometrics from CAC CT. We evaluated whether chamber and myocardial volumes derived from CAC CT improve HF risk prediction beyond traditional factors.

Methods: We analyzed 23,452 adults who underwent CAC CT without prior HF at Houston Methodist. Cardiac chambers (left atrium, right atrium, left and right ventricles) and myocardium were segmented using a validated pre-trained deep learning model (TotalSegmentator). Time-to-incident HF was modeled using Cox proportional hazards models. Model 1 included age, sex, race/ethnicity, and CAC score. Model 2 added chamber and myocardial volumes modeled flexibly using penalized splines. Discrimination was evaluated with Harrell’s C-index, and model improvement assessed by likelihood ratio test.

Results: Over 32,730 person-years of follow-up (median 0.79 years), 330 participants developed HF (incidence rate: 10.1 per 1,000 person-years). Model 2 significantly improved discrimination over Model 1 (C-index: 0.789 vs. 0.706; ΔC = +0.083; p < 2×10^-16) . Subgroup C-indices for Model 2 were: Male 0.806, Female 0.793, White 0.804, Black 0.742, and Hispanic 0.734. Compared to the lowest tertile of predicted risk, the highest tertile had a markedly increased risk of HF (HR: 8.25 [95% CI: 5.47–12.44]; p < 0.001), with a cumulative incidence of 21.8 per 1,000 person-years versus 2.7 in the lowest tertile (Figure).

Conclusion: Cardiac chamber and myocardial volumes extracted from CAC CT significantly enhance HF risk prediction beyond traditional risk factors. The model performs well across sex and racial/ethnic subgroups and supports the potential for equitable, opportunistic HF risk stratification using existing imaging.