Ccta-based AI for Diagnosing ≥ 50 % coronary Stenosis: A patient- and Vessel-Level meta-analysis

To investigate the diagnostic performance of CCTA-based artificial intelligence (AI) in detecting ≥ 50 % coronary stenosis of coronary artery disease (CAD) at both the patient and vessel levels.

A systematic search of PubMed, Embase, and Web of Science databases (from inception to March 2025) was conducted to identify diagnostic studies evaluating CCTA-based AI methods for detecting CAD. Studies of ≥ 50 % coronary stenosis were enrolled. Data for diagnostic performance was extracted and meta analysis was performed. Statistical analyses were performed using RevMan 5.4, Meta-Disc 1.4, and Stata 16.0. This study was designed and reported following the PRISMA-DTA statement.

The systematic search yielded 2,211 potentially relevant records. Following multi-stage screening, 11 eligible studies (1506 patients; 2896 vessels) were included. At the patient level, AI based assistant tools demonstrated pooled sensitivity of 0.95 [95 % CI (0.93-0.97)], specificity of 0.73 [95 % CI (0.61-0.82)], and AUC of 0.96 [95 % CI (0.94-0.97)] for CAD with ≥ 50 % stenosis diagnosis. The positive likelihood ratio (+LR) was 3.5 [95 % CI (2.4-5.2)] and negative likelihood ratio (-LR) was 0.07 [95 % CI (0.05-0.10)], with a pooled DOR of 52 [31-86]. At the vessel level of diagnosing ≥ 50 % stenosis diagnosis, AI-based assistant diagnostic tool showed pooled sensitivity of 0.87 [95 % CI (0.83-0.90)], specificity of 0.89 [95 % CI (0.82-0.93)], +LR of 7.7 [95 % CI (4.8-12.5)], -LR of 0.15 [95 % CI (0.11-0.19)], DOR of 53 [35-79], and AUC of 0.93 [95 % CI (0.90-0.95)].

When pooled across diverse deep-learning systems, AI-assisted CCTA demonstrates high sensitivity and solid diagnostic performance for detecting ≥ 50 % coronary stenosis. However, this reflects aggregated results from heterogeneous models rather than the capability of any single AI tool, limiting direct generalizability to specific systems or vendors.