AI Detects Hidden Coronary Calcium in People Told They Have a Zero Score

For over three decades, a coronary artery calcium (CAC) score of zero has been considered one of the most reassuring findings in preventive cardiology — the so-called 'power of zero.' Yet accumulating data show that a small but meaningful proportion of people with CAC=0 by conventional Agatston scoring (Agatston-1.0) still go on to develop coronary heart disease (CHD). The standard method was designed in the late 1980s using electron-beam CT and relies on a fixed 130 Hounsfield unit (HU) threshold, a minimum lesion area of 1 mm², and 2.5–3 mm slice thickness — parameters that have never been updated despite dramatic improvements in CT technology. These constraints cause partial-volume averaging to obscure small, fragmented, or semi-calcified lesions, leaving real early-stage disease undetected.

To address this gap, the authors developed Agatston-2.0, a fully automated AI framework that replaces fixed thresholds with a continuous, probabilistic, voxel-wise model. The pipeline performs deep-learning–based 3D segmentation of the entire coronary artery tree (left main, LAD, LCX, and RCA), applies scanner-specific intensity calibration using physical hydroxyapatite phantoms or patient-derived reference tissues when phantoms are absent, performs relative density weighting across all voxels, and applies spatial neighborhood filtering to suppress noise while aggregating biologically coherent calcification. The resulting AI-derived CAC score (AI-CAC) is applicable to scans with slice thickness as thin as 0.2 mm and generates a continuous score rather than a binary threshold-dependent output.

The study pooled individual participant-level data from 3,965 adults with a baseline conventional CAC=0: 2,816 from the Multi-Ethnic Study of Atherosclerosis (MESA) and 1,149 from the Framingham Heart Study (FHS). MESA enrolled participants aged 45–84 across six U.S. sites using both electron-beam and multidetector CT; FHS used a 64-detector-row GE scanner at 2.5 mm slice thickness. Both studies employed hydroxyapatite calibration phantoms during scanning. Incident CHD events were tracked for up to 20 years, and associations were modeled using Cox proportional hazards regression adjusted for traditional cardiovascular risk factors including age, sex, race/ethnicity, smoking, diabetes, blood pressure, lipids, and BMI.

The AI-CAC score flagged 862 participants (21.7% of the CAC=0 pool) as having detectable coronary calcification. Over 20 years of follow-up, those with AI-CAC>0 had a CHD incidence of 7.7%, compared with 3.8% in those with AI-CAC=0 (p<0.0001) — roughly a doubling of risk. After full multivariable adjustment, AI-CAC>0 remained independently associated with incident CHD (hazard ratio 1.71, 95% CI 1.18–2.47). Critically, the AI score also strongly predicted conversion from a zero to a positive conventional Agatston score on follow-up imaging (adjusted HR 1.95, 95% CI 1.70–2.24), supporting the interpretation that the AI is detecting real early calcification rather than imaging artifact.

The implications for preventive cardiology are substantial. Clinicians currently use CAC=0 to defer statin therapy and downstream testing. If roughly one in five of those patients actually has occult calcification detectable by AI — and those patients carry nearly twice the 20-year CHD risk — the 'power of zero' guarantee is weaker than previously understood for a meaningful subset. Agatston-2.0 could reframe CAC from a binary risk filter into a continuous, higher-sensitivity biomarker, enabling more precise treatment decisions especially in intermediate-risk individuals. The authors position this as a potential new standard, pending validation in independent cohorts and prospective intervention studies.