Incremental value of a CCTA-derived AI-based ischemia algorithm over standard CCTA interpretation of predicting myocardial ischemia in patients with suspected coronary artery disease.

Background A novel artificial intelligence-guided quantitative computed tomography ischemia algorithm (AI-QCTischemia) comprises a machine-learned method using atherosclerosis and vascular morphology features from coronary computed tomography angiography (CCTA) images to predict myocardial ischemia. This study evaluates the diagnostic performance of AI-QCTischemia compared to standard CCTA interpretation in detecting myocardial ischemia. Methods And Results Patients with suspected coronary artery disease (CAD) undergoing CCTA were analyzed, with ischemia detected by stress [15O]H2O positron emission tomography (PET) as the reference. AI-QCTischemia analysis was successfully completed in 84 ​% of patients undergoing CCTA. A total of 1746 patients (mean age 62 ​± ​10 years, 44 ​% male) were included. In visual CCTA reading, 518 (30 ​%) patients had obstructive CAD, defined as diameter stenosis of ≥50 ​%. Myocardial ischemia on PET was detected in 325 (19 ​%) patients whereas AI-QCTischemia was positive in 430 (25 ​%) patients. The diagnostic accuracy, sensitivity, specificity, positive predictive value, and negative predictive value of the AI-QCTischemia for the assessment of myocardial ischemia were 87 ​%, 81 ​%, 88 ​%, 61 ​%, and 95 ​%, respectively, compared to 86 ​%, 93 ​%, 85 ​%, 58 ​%, and 98 ​% for visual CCTA reading. AI-QCTischemia demonstrated higher diagnostic accuracy, specificity, and positive predictive value, but lower sensitivity and negative predictive value than visual CCTA reading (p-value <0.001). Combining AI-QCTischemia with visual CCTA reading improved ischemia discrimination compared with visual CCTA reading alone (area under the receiver operating characteristic curve 0.899 vs. 0.868, p ​< ​0.001). Conclusions Among patients with suspected CAD, the AI-guided CCTA-derived ischemia algorithm demonstrated improved specificity as compared with visual CCTA reading but this was at a cost of decreased sensitivity, resulting in a slight improvement in diagnostic accuracy for predicting PET-defined myocardial ischemia. These findings suggest that AI-QCTischemia may support clinicians in refining diagnostic decision-making and streamlining patient selection for further testing.