CARDINAL Improves Long-Horizon Risk Prediction From Cardiac CT

Cardiovascular risk tools often depend on clinical variables that may be missing or stale, while non-contrast cardiac CT is commonly reduced to coronary artery calcium or a small set of engineered measurements. The authors argue that this leaves prognostic information in the scan unused. CARDINAL, short for Cardiovascular Assessment via Representation learning from Deep Imaging with Nested Anatomical Latent embeddings, learns a compact CT representation and uses it for major adverse cardiovascular event (MACE) prediction over 1-, 3-, 5-, and 10-year horizons.

Core Contribution

The paper’s central contribution is not simply another CT classifier. It combines imaging representation learning with nested latent embeddings: a single learned representation can be truncated to different dimensions while preserving clinically relevant information. That design lets the authors test whether the predictive signal is concentrated in a small number of latent variables rather than requiring a large opaque feature vector.

The evaluation is also broader than a comparison against calcium alone. CARDINAL is assessed against American Heart Association pooled cohort equations (PCE), PREVENT scores, coronary artery calcium (CAC), segmentation-derived CT biomarkers, and a 70-feature structural-radiomics model. This matters because the practical question is whether end-to-end image representations add information beyond both standard clinical risk equations and hand-built imaging features.

Technical Approach

Patient flow begins with source CT examinations, index-examination selection, cohort eligibility, and a fixed patient-level split into horizon-specific matched held-out test cohorts. Five independently trained models use the same held-out patients, so reported variability reflects model-fold variation rather than a changing test set. The paper distinguishes withFU and ignoreFU evaluation settings, and reports both AUROC and AUPRC across prediction horizons.

The Matryoshka-style representation is the method’s most distinctive component. Rather than optimizing only a full-dimensional embedding, CARDINAL trains nested latent prefixes from 1 through 1,024 dimensions. Figure 2 tracks phenotype recovery at every depth using intraclass correlation coefficients, $R^2$, and mean absolute error in the supplementary material. The setup asks a useful deployment question: how much representation can be discarded before performance meaningfully falls?

Figure 1 makes the evaluation design unusually legible: fixed patient-level held-out cohorts are evaluated by five independently trained fixed-fold models. That is preferable to presenting a single split result, although it does not substitute for an external-site validation cohort.

Results and Analysis

At the 10-year horizon, CARDINAL (joint) reached an AUROC of 0.866 ± 0.020 and an AUPRC of 0.890 ± 0.015. Structural radiomics, identified as the strongest baseline, reached 0.826 ± 0.023 AUROC and 0.826 ± 0.022 AUPRC. The absolute AUROC difference is 0.040, while the AUPRC difference is 0.064. For a long-horizon event-prediction task, the precision-recall gain is the more consequential result: it suggests that the learned CT representation improves ranking where event prevalence makes positive predictions harder to identify.

The authors report that gains are largest at longer horizons. Figure 3 presents mean AUROC and AUPRC at each horizon for both follow-up treatments, alongside PCE, three PREVENT configurations, CAC, and the imaging comparators. It also labels evaluated patient and event counts below each panel. The breadth of these comparisons supports the narrower conclusion that CT-derived learned features can outperform the tested risk-score and handcrafted-imaging alternatives in this cohort; it does not establish that the model will transfer unchanged to other scanners, care systems, or referral populations.

The compactness result is similarly relevant. Figure 4 reports the fraction of peak 10-year withFU AUROC retained as latent depth shrinks, then compares CARDINAL (joint) with the strongest non-joint comparator across sex, age, CAC, and race subgroups. It further shows cumulative MACE incidence for low-, intermediate-, and high-risk tertiles, including number-at-risk tables. The reported survival analysis gives CARDINAL (joint) a C-index of 0.753 ± 0.015 and a high-versus-low tertile hazard ratio of 10.78 ± 3.16. Those are clinically interpretable stratification results, but calibration is described as exploratory, so the evidence is stronger for discrimination than for assigning an absolute individual risk.

Limits in Practice

The evidence comes from one 17,659-patient cohort and a held-out internal test design. Five fixed-fold models improve confidence in the reported estimates, but they do not test geographic or temporal transport. The paper compares several meaningful baselines, yet its stated findings should be read as comparative performance within this study’s cohort and endpoint definitions. Before use in clinical decision-making, external validation and fuller calibration assessment would be needed, particularly because risk-tertile separation does not by itself show that predicted probabilities are well calibrated.