Hybrid deep learning time-to-event modeling of major adverse cardiovascular events using coronary artery calcium score scans.

BACKGROUND: The Agatston score from non-contrast coronary calcium scoring scan (CACS) is a robust predictor of major adverse cardiovascular events (MACE) but summarizes total calcium burden and may miss other risk-related imaging information in CACS. We hypothesized that a hybrid deep-learning (DL) model integrating image features with clinical, Agatston, and calcium-omics data could improve MACE risk prediction. METHODS: In this retrospective cohort of 1950 patients, a DenseNet-121 encoder extracted DL-derived features from CACS volume, which were fused with tabular data (clinical features, Agatston, and 15 calcium-omics features) within a deep survival model. Performance was assessed via five-fold cross-validation using Harrell's concordance-index (C-index) and time-dependent area under the curve (AUC), and net reclassification index (NRI), relative to tabular-only baseline. RESULTS: Expanding the tabular feature set from clinical variables alone to include Agatston and calcium-omics improved discrimination for both tabular-only baseline and hybrid models. Incorporating DL-derived imaging representations consistently outperformed tabular baselines in C-indices (0.802 vs. 0.789 with the full feature set) and yielded higher time-dependent AUCs during first three years of follow-up. The hybrid model also achieved an NRI of 0.119 (p = 0.037). Patients reclassified from low to high risk had a significantly higher event rate (37.5%) compared to those remaining low risk (1.1%, p < 0.001). CONCLUSION: These preliminary findings suggest that integration of DL-derived CACS representations with clinical and hand-crafted features can provide complementary prognostic information and modestly improve MACE risk stratification compared to tabular data alone.