Fully Automated, Explainable AI Analysis of Coronary CTA Improves Prediction of Major Cardiovascular Events.
Authors
Affiliations (3)
Affiliations (3)
- Siemens Healthineers, Princeton, New Jersey, USA.
- Valley Health System, Paramus, New Jersey, USA.
- Valley Health System, Paramus, New Jersey, USA. Electronic address: [email protected].
Abstract
Coronary artery disease (CAD) remains a major global cause of morbidity and mortality. Coronary computed tomography angiography (CCTA) is central to noninvasive diagnosis and risk assessment, yet current workflows are labor-intensive and depend on semiquantitative scoring systems that do not fully leverage prognostic information in CCTA. The objectives of the study were to develop and validate a fully automated, explainable artificial intelligence (AI) system (AI-automated coronary analysis [ACA]) for comprehensive CCTA analysis to predict major adverse cardiovascular events and compare its performance with conventional risk markers. AI-ACA automatically extracts a suite of CCTA-derived biomarkers, including stenosis severity, plaque burden, and composition; AI-derived CAD - reporting and data system (CAD-RADS) (AI-CADRADS); coronary tortuosity; and subtended myocardial mass, to generate interpretable risk estimates. In a retrospective cohort (n = 1,702), we compared AI-ACA with Agatston score, standard demographic risk markers, and AI-CADRADS. Model explainability was assessed using Shapley Additive Explanations. AI-ACA demonstrated higher predictive performance (area under the curve [AUC]: 0.81) than Agatston score (AUC: 0.76) and demographic markers (AUC: 0.67), and similar performance to AI-CADRADS (AUC: 0.81). AI-ACA showed better prognostic discrimination than Agatston score, reflected by a higher Harrell C-index (0.791 vs 0.742) and greater separation in event-free survival on Kaplan-Meier analysis (log-rank 31.79 vs 13.71). Shapley Additive Explanations analysis identified total plaque volume and burden, maximum stenosis severity, and AI-CADRADS as the strongest contributors to risk estimation. A fully automated and explainable AI platform for CCTA enables comprehensive quantitative characterization of CAD without manual intervention. Although prognostic performance was comparable to AI-CADRADS and modestly higher than Agatston score, the platform provides transparent biomarker-based risk assessment and scalable analysis that may facilitate broader clinical implementation of quantitative CCTA.