Agatston 2.0: AI-derived calcium burden and plaque density profiling improve coronary heart disease risk stratification within CAC scores 1–99
A Azimi, K Atlas, C Zhang, A Reeves, R Mirjalili, M MozafaryBazargany, A Hashemi, A Ghaffari Jolfayi, H Mohammadi, H Zarei, H Ghoshouni, F Fayad, D Yankelevitz, N Wong, M NaghaviAbstract
Background
Coronary artery calcium (CAC) scoring is recommended to refine cardiovascular risk in primary prevention, with scores 1–99 generally associated with low-to-intermediate risk. However, outcomes within this range are heterogeneous and not fully explained by Agatston scoring. Conventional Agatston scoring uses a fixed ≥130 HU threshold for plaque detection, potentially overlooking lower-attenuation calcification, and its peak-density weighting does not yield a continuous plaque density profile.
Purpose
To develop and validate Agatston 2.0, an AI-based framework integrating continuous calcium burden quantification and plaque density profiling, and to evaluate its incremental prognostic value beyond Agatston scoring among individuals with CAC 1–99.
Methods
We analysed 1,238 participants from a prospective multi-ethnic cohort with baseline Agatston CAC 1–99, followed for incident CHD (median 17.4; maximum 19.4 years). Agatston 2.0 generates two outputs: (1) a continuous AI-derived calcium score using probabilistic voxel-wise quantification without fixed attenuation thresholds, capturing subthreshold burden; and (2) a plaque density metric quantifying hyperdense voxels exceeding 10 standard deviations above scanner calibration, isolating mineralisation independently of total volume. Participants were classified into four strata: AI-derived score ≤99 or >99 crossed with plaque density top quartile (dense) versus lower three quartiles (non-dense); the AI-derived ≤99 non-dense group served as reference. Cox models were adjusted for age, sex, race/ethnicity, blood pressure, antihypertensive use, LDL cholesterol, HDL cholesterol, triglycerides, lipid-lowering therapy, diabetes, and smoking. Discrimination was assessed using C-index and net reclassification improvement (NRI).
Results
Among 1,238 participants (153 CHD events; 12.4%; mean age 64.4±9.5 years; 53.2% male), an AI-derived score >99 versus ≤99 was associated with higher CHD risk (HR 1.54, 95%CI 1.04–2.30), attenuating after full adjustment (HR 1.37, 0.92–2.06). Higher plaque density (top versus bottom quartile) was inversely associated with CHD (HR 0.55, 0.34–0.88; adjusted HR 0.58, 0.35–0.95). Kaplan–Meier analysis demonstrated a risk gradient across strata: 19-year cumulative incidence was 26.7% in the >99 non-dense group (log-rank p=0.0095), 13.7% in >99 dense (p=0.8533), 15.3% in ≤99 non-dense (reference), and 9.4% in ≤99 dense (p=0.0420). Adding Agatston 2.0 to Agatston scoring significantly improved discrimination (C-index 0.619 [0.572–0.660] vs 0.558 [0.507–0.609]; p=0.006) and reclassification (overall NRI +0.247, 95%CI +0.107 to +0.389; p<0.001).
Conclusion
Agatston 2.0 improves discrimination and reclassification within CAC 1–99 by integrating calcium burden with plaque density profiling. An AI-derived score >99 combined with lower density identifies individuals at higher long-term CHD risk who may not be distinguished by Agatston scoring alone.Cummulative incidence3D risk plot