Diagnostic Performance of AI-Assisted Coronary CT Angiography: A Systematic Review and Meta-Analysis

Coronary computed tomography angiography (CCTA) is vital for diagnosing ischemic heart disease, yet its accuracy is unreliable due to varying reader expertise. Artificial Intelligence (AI)-driven automated stenosis assessment offers promise for enhancing diagnostic consistency. We aim to evaluate an AI-CCTA assessment against invasive coronary angiography, invasive FFR, and expert readings.

We performed a comprehensive search in Web of Science, Scopus, PubMed, Cochrane Library, and EMBASE from inception until March 2025. Two independent reviewers screened articles and extracted data on study design, patient demographics, AI methodology, stenosis thresholds, and outcomes. For statistical analysis, we constructed summary receiver operating characteristic (SROC) curves and used a bivariate random-effects model to derive pooled sensitivity, specificity, diagnostic odds ratios (DOR), and area under the curve (AUC). Forest plots were generated to visualize these metrics.

Our meta-analysis included 34 studies with 10,067 patients. AI-based CCTA demonstrated excellent diagnostic performance with an AUC of 0.932 for per-patient analysis. The pooled per-patient sensitivity was 0.89 (95 % CI: 0.87-0.91) and specificity was 0.80 (95 % CI: 0.74-0.86) with diagnostic OR of 37.07 (95 % CI: 24.57-55.92). AI validated against expert readers achieved the highest accuracy (0.94, 95 % CI: 0.87-0.98). The > 70 % stenosis threshold demonstrated superior performance (accuracy: 0.90, specificity: 0.96) compared to the > 50 % threshold (accuracy: 0.85, specificity: 0.87). Per-vessel analysis showed comparable results with an AUC of 0.905.

AI-assisted coronary CT angiography delivers high diagnostic performance for coronary stenosis detection, with strong AUC values, high sensitivity and specificity, and robust diagnostic OR across both per-patient and per-vessel assessments.