AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Prognostic Value of Artificial Intelligence Based Coronary Artery Calcium Scoring on Non-Gated Chest Computed Tomography Scan for Predicting Major Adverse Cardiovascular Events - A Systematic Review and Meta-Analysis
By SHAH D, Shah I, Davies D et al.
Background: Artificial intelligence (AI) based coronary artery calcium (CAC) quantification enables automated cardiovascular risk assessment from routine non-gated chest computed tomography (CT) without dedicated cardiac imaging or manual interpretation. While prior studies demonstrated diagnostic feasibility, the pooled prognostic significance of AI-derived CAC for predicting major adverse cardiovascular events (MACE) remains unclear. Methods: We systematically searched PubMed and Cochrane Library through January 2026, supplemented by bibliography screening (PROSPERO: CRD420261406423). Inclusion criteria were studies assessing machine learning or deep learning enabled CAC scoring on non-gated chest CT with longitudinal MACE outcomes. The primary outcome was the adjusted hazard ratio (aHR) for MACE comparing severe CAC burden (CAC>400) versus CAC 0. Secondary outcomes included net reclassification improvement (NRI), C-index discrimination, and concordance with expert readers. Risk of bias was assessed using PROBAST. Pooled estimates were derived using the DerSimonian-Laird random effects model. Sensitivity, influence, and publication bias analyses examined robustness and heterogeneity. Results: 12 studies met inclusion criteria, 8 provided aHR suitable for quantitative pooling (4 contributed only to secondary outcomes due to absence of extractable aHR for MACE). The 8 pooled studies comprised 58,450 participants with follow-up of 2.4 to 6.7 years (median 4.2). PROBAST classified 6 of 8 studies as low risk of bias. Severe AI-derived CAC burden (CAC>400) was independently associated with MACE compared with CAC 0 (pooled aHR 2.95, 95% CI 2.64-3.29; I ² =53%). Influence analysis identified Pieszko et al. (2022) as an outlier, and its exclusion strengthened the association while resolving heterogeneity (aHR 3.34, 95% CI 2.89-3.86, I ² =21.8%). Restriction to PROBAST low-risk studies yielded consistent estimates (aHR 2.99, 95% CI 2.67-3.36). No significant publication bias was detected (Egger’s p=0.82). Integration of AI-CAC into traditional clinical risk models improved reclassification (pooled NRI 0.16, 95% CI 0.07-0.25), though between-study heterogeneity was high (I ² =80%). Discrimination was strong across studies (C-index 0.7-0.85), and AI-derived CAC showed substantial agreement with expert reader assessment (weighted k 0.60-0.95). Conclusions: AI-based CAC scoring from routine non-gated chest CT is strongly associated with future MACE and improves cardiovascular risk prediction beyond traditional clinical risk factors. Given the millions of non-gated chest CTs performed annually for non-cardiac indications, automated AI-CAC scoring represents a scalable, no-additional-cost opportunity for opportunistic cardiovascular risk stratification and identification of high-risk patients who would otherwise go unrecognized.