Article Impact Level: HIGH Data Quality: STRONG Summary of NEJM AI, 2(6). https://doi.org/10.1056/AIoa2400937 Dr. Raffi Hagopian et al.
Points
- Researchers developed a deep learning algorithm to automatically quantify coronary artery calcium from millions of routinely performed, non-gated chest CT scans for opportunistic cardiovascular risk screening.
- The AI-CAC model accurately differentiated between zero and non-zero calcium scores with 89.4% accuracy and identified high-risk scores exceeding 100 with 87.3% accuracy, compared to standard gated scans.
- Patients identified with a high calcium score over 400 had a significantly increased risk of 10-year all-cause mortality, with a hazard ratio of 3.49 against the zero-score group.
- The model was also highly predictive of major adverse events, showing a threefold increased risk of stroke, myocardial infarction, or death for patients in the highest calcium score category.
- In a simulated screening of over 8,000 individuals, cardiologists confirmed that 99.2% of patients identified by the AI with very high scores would benefit from lipid-lowering therapy.
Summary
Researchers have developed and validated a deep learning algorithm, AI-CAC, for the automated quantification of coronary artery calcium (CAC) from routine non-gated chest computed tomography (CT) scans. The model was trained on 446 expert segmentations from a heterogeneous dataset of imaging studies conducted across 98 medical centers within the U.S. Department of Veterans Affairs national healthcare system. The algorithm’s performance was evaluated against the clinical standard of electrocardiogram (ECG)-gated CAC scoring in a cohort of 795 patients who had both gated and non-gated scans performed within one year of each other. A further test to simulate opportunistic screening was conducted on 8,052 low-dose CT (LDCT) scans.
The AI-CAC model demonstrated strong diagnostic and prognostic capabilities. When compared to gated scans, the algorithm differentiated between zero and nonzero Agatston scores on non-gated CTs with an accuracy of 89.4% (F1 score 0.93) and distinguished scores of less than 100 from those of 100 or greater with an accuracy of 87.3% (F1 score 0.89). Opportunistically derived AI-CAC scores were highly predictive of clinical outcomes. Over a 10-year follow-up, patients with a CAC score of 0 had a 25.4% all-cause mortality rate, compared to 60.2% for those with a score over 400 (Cox Hazard Ratio [HR] 3.49; P<0.005). Similarly, for the composite endpoint of first-time stroke, myocardial infarction, or death, event rates were 33.5% for the CAC 0 group versus 63.8% for the >400 group (HR 3.00; P<0.005).
In the simulated screening cohort of 8,052 LDCTs, the AI-CAC model identified 3,091 individuals (38.4%) with CAC scores greater than 400. A qualitative review by four cardiologists of a random sample of 531 of these high-risk patients confirmed that 527 (99.2%) would be candidates for lipid-lowering therapy based on their CAC burden. These findings suggest that the AI-CAC tool can effectively leverage millions of existing non-gated chest CTs for opportunistic cardiovascular risk stratification. While developed on a veteran population, the publicly available model shows significant potential to shift clinical practice towards more proactive cardiovascular disease prevention.
Link to the article: https://ai.nejm.org/doi/10.1056/AIoa2400937
References Hagopian, R., Strebel, T., Bernatz, S., Myers, G. A., Offerman, E., Zuniga, E., Kim, C. Y., Ng, A. T., Iwaz, J. A., Nürnberg, L., Singh, S. P., Carey, E. P., Kim, M. J., Schaefer, R. S., Yu, J., Gentili, A., & Aerts, H. J. W. L. (2025). Ai opportunistic coronary calcium screening at veterans affairs hospitals. NEJM AI, 2(6). https://doi.org/10.1056/AIoa2400937
