Prognostic Value of AI-Based Quantitative Coronary CTA vs Human Reader-Based Visual Assessment

BACKGROUND: The severity and extent of whole heart coronary plaque volume and stenosis can be reliably measured by artificial intelligence-guided quantitative coronary computed tomography angiography (AI-QCT). Limited data are available on the potential incremental prognostic value compared with currently recommended qualitative coronary computed tomography angiography (CTA) reads and the coronary artery calcium score (CACS). OBJECTIVES: The aim of this study was to evaluate the prognostic value of AI-QCT compared with human coronary CTA reads, including the CAD-RADS (Coronary Artery Disease-Reporting and Data System), CACS, and the modified Duke Index. METHODS: CONFIRM2 (Quantitative COroNary CT Angiography Evaluation For Evaluation of Clinical Outcomes: An InteRnational, Multicenter Registry) is a multicenter, international, observational cohort study of patients undergoing clinically indicated coronary CTA with follow-up for major adverse cardiac events (MACE). Asymptomatic patients and those with cardiac history were excluded. Coronary artery disease presence, extent, and composition were quantified by AI-QCT across the coronary tree, yielding 24 patient-, vessel-, and plaque-level variables. On the basis of prior analyses, noncalcified plaque burden and diameter stenosis were identified as the strongest predictors and combined for statistical modeling as "AI-QCT." Comparator computed tomography scores included CAD-RADS, CACS, and the modified Duke Index, whereas clinical predictors were summarized in the risk factor-weighted clinical likelihood score. Area under the curve (AUC) and continuous net reclassification index (NRI) were calculated to assess the incremental value. The primary endpoint was MACE (death, myocardial infarction [MI], stroke, heart failure, late revascularization, or hospital stay for unstable angina), and the secondary endpoint was death or MI. RESULTS: In 1,916 patients with all risk scores available, 87 (4.5%) MACE and 27 (1.4%) death/MI events occurred during 3 years of follow-up. There was a stepwise risk increase with higher coronary artery disease classifications with CAD-RADS and CACS. The addition of AI-QCT significantly improved risk stratification for MACE compared with CAD-RADS (AUC: 0.81 vs 0.79; P < 0.001 and NRI: 0.47; P < 0.001), CACS (AUC: 0.79 vs 0.70; P < 0.001 and NRI 0.61; P < 0.001), the modified Duke Index (AUC: 0.81 vs 0.76; P < 0.001 and NRI: 0.52; P < 0.001), and CAD-RADS + CACS model (AUC: 0.81 vs 0.79; P = 0.004 and NRI: 0.54; P < 0.001). AI-QCT also improved discrimination when results were adjusted for the risk factor-weighted clinical likelihood and for the prediction of death/MI. Excluding 195 patients with severe stenosis (≥70%), in a multivariable model of CAD-RADS and AI-QCT, only AI-QCT was significantly associated with MACE and death/MI, and AI-QCT significantly improved risk stratification compared with CAD-RADS for MACE (AUC: 0.77 vs 0.72; P < 0.001 and NRI: 0.54; P < 0.001) and death/MI (AUC: 0.81 vs 0.73; P = 0.011 and NRI: 0.69; P = 0.001). CONCLUSIONS: AI-QCT provided incremental prognostic information compared with CAD-RADS 2.0, CACS, and the modified Duke Index for the prediction of MACE as well as the secondary endpoint of death or nonfatal MI.