Diagnostic accuracy of artificial intelligence–enhanced coronary CT angiography for detecting functionally significant coronary artery disease: A systematic review and meta-analysis

BACKGROUND: Coronary computed tomography angiography (CCTA) is widely used to evaluate suspected coronary artery disease (CAD), but its ability to determine the functional significance of coronary stenoses remains limited. Artificial intelligence (AI)-enhanced CCTA has emerged as a promising noninvasive approach to improve ischemia assessment. OBJECTIVES: To evaluate the diagnostic accuracy of AI-based CCTA for detecting hemodynamically significant CAD using invasive reference standards. METHODS: We conducted a systematic review and diagnostic test accuracy meta-analysis in accordance with PRISMA-DTA guidelines. Studies evaluating AI algorithms applied to CCTA for the detection of functionally significant CAD were eligible if invasive fractional flow reserve (FFR) or invasive coronary angiography served as the reference standard. Risk of bias was assessed using QUADAS-2. The primary analysis was restricted to studies using FFR ≤ 0.80. Diagnostic performance was estimated using a bivariate random-effects hierarchical summary receiver operating characteristic (HSROC) model. Prespecified sensitivity analyses evaluated the effects of alternative reference standards, study quality, verification strategy, AI methodology, and unit of analysis. RESULTS: Thirty-five studies involving approximately 8400 participants met the eligibility criteria, of which 18 contributed to the quantitative synthesis. For studies using FFR ≤ 0.80 as the reference standard (13 studies), pooled sensitivity was 0.823 (95% CI, 0.761-0.872) and pooled specificity was 0.820 (95% CI, 0.732-0.883). The bivariate HSROC model produced similar estimates (sensitivity 0.827; specificity 0.820). Sensitivity analyses excluding studies at high risk of bias and including studies using alternative physiological reference standards (iFR ≤ 0.89 or FFR-based composite definitions) demonstrated comparable diagnostic performance. Exploratory subgroup analyses showed generally consistent accuracy across AI methodologies, verification strategies, and units of analysis, although heterogeneity remained. CONCLUSIONS: AI-enhanced CCTA demonstrates good and balanced diagnostic accuracy for identifying functionally significant CAD compared with invasive reference standards. Diagnostic performance remained robust across multiple sensitivity analyses, supporting the potential role of AI-assisted CCTA as a noninvasive gatekeeper to invasive coronary angiography. Further prospective multicenter studies using standardized AI algorithms and external validation are needed before widespread clinical implementation.