Introduction
Opportunistic detection of coronary artery calcium (CAC) on routine, non-contrast chest CT is a scalable strategy for cardiovascular risk stratification — it uses scans that have already been performed for other reasons. But a practical question stands in the way of wide adoption: how consistent is automated CAC detection when it moves out of a single research setting and into heterogeneous, real-world clinical environments?
This study, presented at the 2026 Society of Cardiovascular Computed Tomography (SCCT) Annual Meeting by investigators from Mass General Brigham, University of Washington and Wake Forest Health, evaluated the agreement between Nanox.AI’s FDA-cleared coronary calcium algorithm (HealthCCSng) and expert visual assessment across three independent clinical sites. This page summarizes the design, headline results, and why cross-site consistency matters, and links to the full poster.
Key Takeaways
- The analysis compared an FDA-cleared AI CAC algorithm (Nanox.AI HealthCCSng) against expert visual assessment by cardiac imaging teams.
- It spanned 3 independent clinical sites and 5,468 non-contrast chest CT studies.
- Overall agreement between AI and expert readers exceeded 90% across all CAC categories.
- For scans with no calcium (CAC = 0), agreement was 99.0% at Site 1 and 96.2% at Site 3 — high concordance with expert readers on ruling calcium in or out.
- Consistent agreement across sites supports the use of opportunistic, AI-enabled CAC detection in real-world settings.
Study Overview
The study set out to evaluate agreement between an FDA-cleared AI cardiac solution (Nanox.AI HealthCCSng) and expert visual assessment across multiple clinical sites, using clinically relevant CAC thresholds. Rather than measuring accuracy against a reference standard in a single controlled dataset, it tested how consistently the AI matched expert readers across heterogeneous, real-world imaging environments.
- Comparison: AI algorithm (automated scoring) versus expert reader (visual assessment).
- Sites: 3 independent clinical sites.
- Sample: 5,468 non-contrast chest CT studies.
- AI solution: Nanox.AI HealthCCSng (FDA-cleared).
- Framework: Agreement evaluated by site and by CAC category using clinically relevant thresholds.
Key Findings
Across the pooled analysis, AI-based CAC detection showed high, consistent agreement with expert interpretation.
- Overall agreement exceeded 90% across all CAC categories.
- Agreement was reported by site and by CAC category, with performance color-graded (85–90%, 90–95%, ≥95%).
- For the CAC = 0 category, agreement was 99.0% at Site 1 and 96.2% at Site 3 — indicating high, consistent agreement with expert visual assessment on scans with no detectable coronary calcium. CAC = 0 agreement is reported for two sites because the third site’s data were collected before the CAC = 0 category was introduced.
High agreement in the CAC = 0 category means the AI and the expert readers concurred on which scans showed no detectable calcium, while agreement in the higher categories means they concurred on where calcium was present.
Note: These are results presented in abstract/poster form (SCCT 2026, Abstract 565; abstract page: https://cattendee.abstractsonline.com/meeting/21508/presentation/578). Category- and site-level figures beyond those shown here should be confirmed against the full poster and any peer-reviewed publication.
Why Cross-Site Consistency Matters
An AI tool is only useful for population-scale prevention if it performs consistently outside the lab.
- Real-world evaluation — testing across three independent sites reflects the variability of everyday clinical practice, not a single curated dataset.
- Consistency builds trust — high, stable agreement with expert readers supports clinician confidence in automated CAC scoring.
- Enables opportunistic screening — consistent AI analysis of routine CT scans lets health systems extract cardiovascular risk information from imaging already being performed.
- Supports preventive pathways — consistent CAC identification is the first step toward acting on incidental findings, as explored in the companion AI INFORM trial on preventive therapy.
Frequently Asked Questions
It measured the agreement between an FDA-cleared AI coronary artery calcium algorithm and expert visual assessment by cardiac imaging teams across three clinical sites and 5,468 non-contrast chest CT studies.
Overall agreement exceeded 90% across all CAC categories. For scans with no calcium (CAC = 0), agreement reached 99.0% at one site and 96.2% at another; the third site’s data predate the CAC = 0 category, so no CAC = 0 figure is reported for it.
Performance can vary with scanner type, protocol, and patient population. Consistent agreement across independent sites indicates that the AI agrees with expert readers in diverse real-world settings, not just in a single dataset.
It is the practice of detecting coronary artery calcium on CT scans performed for non-cardiac reasons, allowing cardiovascular risk to be assessed without additional imaging or radiation.
Client portal