Agatston 2.0: a next-generation AI-based coronary calcium quantification approach to improve risk stratification among individuals with zero Agatston scores
A Azimi, K Atlas, C Zhang, A Reeves, R Mirjalili, M MozafaryBazargany, A Ghaffari Jolfayi, A Hashemi, H Ghoshouni, H Mohammadi, H Zarei, Z Fayad, D Yankelevitz, N Wong, M NaghaviAbstract
Background
Coronary artery calcium (CAC) scoring using the Agatston method is widely applied to detect subclinical coronary artery disease. A CAC score of zero confers low cardiovascular risk, known as the “Power of Zero.” However, a subset of individuals with zero Agatston scores still experience coronary heart disease (CHD) events, underscoring limitations of conventional threshold-based scoring.
Objectives
To develop and validate an artificial intelligence–based CAC quantification framework (Agatston 2.0) to improve detection of subclinical disease and long-term risk stratification among individuals with baseline Agatston CAC=0.
Methods
We pooled 3,965 participants with baseline Agatston CAC=0 from the Multi-Ethnic Study of Atherosclerosis (MESA, n=2,816) and the Framingham Heart Study (FHS, n=1,149). Median follow-up in the pooled cohort was 16.2 years, with maximum follow-up of 19.8 years. Agatston 2.0 applied automated deep learning–based coronary artery segmentation and continuous probabilistic calcium quantification without fixed attenuation or size thresholds to generate a continuous AI-derived CAC score (AI-CAC). Outcomes included myocardial infarction (MI), hard CHD (MI, resuscitated cardiac arrest, CHD death), all CHD (hard CHD plus angina with revascularization), and Agatston score progression, defined as conversion to CAC >0 on follow-up CT imaging during MESA Exams 2–5. Multivariable Cox models adjusted for demographics, cardiovascular risk factors, medication use, and cohort.
Results
Agatston 2.0 detected coronary calcification in 862 participants (21.7%) initially classified as CAC=0 by conventional scoring. At 20 years, cumulative incidence of MI was 4.5% in AI-CAC>0 versus 2.4% in AI-CAC=0 (p=0.0017), hard CHD was 6.4% versus 3.2% (p<0.0001), and all CHD was 7.7% versus 3.8% (p<0.0001). In fully adjusted models, AI-CAC>0 independently predicted MI (HR 1.73, 95% CI 1.06–2.82), hard CHD (HR 1.85, 95% CI 1.23–2.78), and all CHD (HR 1.71, 95% CI 1.18–2.47). AI-CAC>0 also strongly predicted Agatston progression (66.3% vs 43.4% at 10 years; fully adjusted HR 1.95, 95% CI 1.70–2.24; p<0.0001). Among AI-CAC>0 participants, the highest sex-specific tertile demonstrated the greatest risk for all CHD (fully adjusted HR 1.85, 95% CI 1.12–3.07), hard CHD (fully adjusted HR 1.88, 95% CI 1.06–3.32), and Agatston progression (fully adjusted HR 1.87, 95% CI 1.52–2.29). Each 1-SD increase in AI-CAC was associated with an 8–13% higher risk across outcomes. Scan–rescan reproducibility was excellent (r=0.989).
Conclusions
Agatston 2.0 identifies clinically meaningful subclinical coronary calcification in approximately one in five individuals with conventional CAC=0. The AI-derived score provides independent and durable long-term risk stratification and predicts future calcium progression, refining the biological interpretation of the “Power of Zero.”Method for Agatson 2.0KM analysis