Highlights
- AI-CAC, a deep learning algorithm, accurately quantifies coronary artery calcium (CAC) on noncontrast, nongated CT scans from diverse clinical settings.
- Performance of AI-CAC rivals standard ECG-gated CAC scoring and is strongly predictive of all-cause mortality and major cardiovascular events.
- Opportunistic screening using AI-CAC from low-dose chest CTs identifies a significant proportion of patients who could benefit from preventive therapy.
- Model and code are openly available, promoting reproducibility and clinical translation in large health care systems.
Study Background and Disease Burden
Coronary artery calcium (CAC) scoring is one of the most powerful predictors of future cardiovascular events, providing incremental risk stratification beyond traditional clinical factors. However, CAC assessment is typically limited to dedicated, ECG-gated cardiac CT scans. In the United States, millions of chest CT scans—often performed for other indications such as lung cancer screening or pulmonary disease—contain valuable, yet untapped, information about coronary calcification. This gap leaves a substantial portion of the population with unrecognized and untreated atherosclerotic risk, representing a critical missed opportunity for early intervention, especially in high-risk groups such as veterans who often exhibit a higher burden of cardiovascular disease.
Study Design
This study, recently published in NEJM AI, describes the development and validation of AI-CAC, a deep learning algorithm trained to automatically quantify CAC from noncontrast, nongated CT scans. Key distinguishing features include:
- Population and Setting: Imaging data derived from 98 Veterans Affairs (VA) medical centers nationwide, ensuring robust representation of scanner types, imaging protocols, and patient demographics.
- Training and Validation: The algorithm was trained on 446 expert-annotated segmentations. Validation was performed using 795 patients with paired ECG-gated and nongated scans (within 1 year), providing a rigorous head-to-head comparison with the clinical gold standard.
- Opportunistic Screening Simulation: The model was then applied to 8,052 low-dose CTs (LDCTs) to mimic real-world opportunistic CAC screening among veterans undergoing imaging for noncardiac reasons.
- Endpoints: Diagnostic accuracy (zero vs nonzero, <100 vs ≥100 Agatston score), prediction of 10-year all-cause mortality and composite cardiovascular outcomes, and clinical actionability (potential benefit from lipid-lowering therapy as determined by cardiologist review).
Key Findings
Diagnostic Performance
- AI-CAC distinguished zero vs nonzero Agatston scores with 89.4% accuracy (F1 score 0.93).
- It differentiated <100 vs ≥100 Agatston score with 87.3% accuracy (F1 score 0.89).
- These results closely approximate those of standard ECG-gated CAC scoring, demonstrating high reliability of the AI approach across heterogeneous, real-world VA imaging data.
Prognostic Value
- AI-CAC–derived scores from nongated scans were strongly predictive of 10-year all-cause mortality (CAC 0 vs >400: 25.4% vs 60.2%, hazard ratio 3.49, P<0.005).
- Similarly, for the composite outcome of first-time stroke, myocardial infarction, or death, the hazard ratio was 3.00 (P<0.005) comparing CAC 0 vs >400 groups.
Opportunistic Screening Yields
- Among 8,052 veterans undergoing LDCT, 3,091 (38.4%) had AI-CAC scores >400—a group at substantial risk for future cardiovascular events.
- Cardiologist review of a random sample of these high-risk cases confirmed that 99.2% would be eligible for lipid-lowering therapy according to current preventive guidelines, highlighting the clinical actionability of AI-CAC–identified cases.
Open Science and Implementation
- The full AI-CAC model code and weights have been made publicly available, fostering transparency, reproducibility, and potential for rapid dissemination across other health systems.
Expert Commentary
The integration of artificial intelligence into routine clinical imaging workflows represents a paradigm shift in preventive cardiology. The VA study uniquely addresses key barriers to CAC screening by leveraging the vast, underutilized reservoir of chest CTs performed for noncardiac indications. The high diagnostic concordance with ECG-gated scans and strong prognostic associations suggest that AI-CAC can identify high-risk individuals who might otherwise be missed by standard screening protocols.
Several strengths distinguish this study:
- Generalisability: The use of multicenter, multi-scanner data with broad patient heterogeneity enhances real-world applicability.
- Clinical Impact: The ability to prompt guideline-recommended preventive therapy in nearly all high-CAC patients exemplifies the translational value of opportunistic screening.
- Scalability: Automated AI algorithms can be deployed at scale with minimal additional cost or workflow disruption.
However, limitations must be acknowledged:
- While the algorithm performed robustly across VA centers, external validation in other health systems and patient populations is needed to confirm generalizability.
- False positives and negatives, though uncommon, could have downstream implications for patient management and resource utilization.
- Integration of AI-CAC into clinical decision support tools and EHRs will require careful consideration of workflow, provider education, and patient communication.
Conclusion
The VA’s AI-CAC algorithm demonstrates that automated CAC quantification from routine, nongated chest CTs is accurate, actionable, and prognostically meaningful. This approach has the potential to transform opportunistic cardiovascular risk assessment, especially in high-burden populations where dedicated cardiac imaging may not be feasible. Widespread adoption could lead to earlier identification and treatment of at-risk patients, narrowing the gap in preventive cardiovascular care. Future research should focus on prospective implementation, health economic analyses, and broader validation to fully realize the promise of AI-driven population health screening.
References
- Hagopian R, et al. AI Opportunistic Coronary Calcium Screening at Veterans Affairs Hospitals. NEJM AI. 2025;2(6). DOI: 10.1056/AIoa2400937
- Blaha MJ, et al. Coronary artery calcium scoring: Is it time for universal screening? J Am Coll Cardiol. 2017;70(21):2702-2716. doi:10.1016/j.jacc.2017.10.010
- Hecht HS, et al. 2018 AHA/ACC guideline on the management of blood cholesterol: A report of the American College of Cardiology/American Heart Association Task Force on Clinical Practice Guidelines. Circulation. 2019;139(25):e1082-e1143.