Of 100 prospectively recruited asymptomatic male marathon runners 94 CCTA scans were included for CAD analysis (Fig. 1). Six subjects were excluded, five due to severe imaging artifacts that inhibited evaluation of coronary arteries, one due to an unenhanced scan.

Flowchart of subject inclusion. 100 asymptomatic male marathon runners were recruited for CCTA. Six subjects were excluded due to artifacts that severely impaired evaluation of coronary arteries (n = 5) and one subject due to an unenhanced scan (n = 1). 94 subjects were included for final analysis of coronary artery disease and fractional flow reserve

Detailed characteristics of the 94 included subjects are provided in Table 1.

Quantitative variables are expressed as means ± standard deviations followed by ranges in brackets. Qualitative variables are expressed as raw numbers; numbers in parentheses are percentages.

CAD: Coronary Artery Disease; CCTA: coronary computed tomography angiography.

Details of coronary artery stenosis evaluation are provided in Table 2. The workflow of this study is illustrated in Fig. 2. An example of artificial intelligence-based CAD analysis is provided in Fig. 3.

Illustration of coronary artery disease analysis workflow in the study. Coronary CT angiography images were visually interpreted and coronary stenosis was classified according to CAD-RADS (0–5). Curved multiplanar reformations were processed by artificial intelligence-based models (CorEx, Spimed) on a local server. Stenosis was classified according to CAD-RADS (0–5) and FFR was interpreted regarding hemodynamically significant CAD (threshold 0.8)

Example of artificial intelligence-based analysis of coronary artery disease in a subject with CAD-RADS 4 stenosis. A Coronary artery stenosis was automatically quantified using curved multiplanar reformations. In this example there was a significant stenosis of the LAD (> 70%, CAD-RADS 4). B Results of the artificial intelligence evaluation indicated high confidence of the analysis. An additional model calculated fractional flow reserve < 0.8

Variables are expressed as raw numbers; numbers in parentheses are percentages.

CAD-RADS: Coronary Artery Disease-Reporting and Data System 2.0; CCTA: coronary computed tomography angiography; CX: circumflex artery; LAD: left anterior descending artery; RCA: right coronary artery.

In visual analysis, on subject level stenosis > 0% in at least one coronary artery was present in 34/94 (36.2%) subjects with a severity of CAD-RADS 1 in 19/94 (20.2%), CAD-RADS 2 in 12/94 (12.8%), CAD-RADS 3 in 2/94 (2.1%), CAD-RADS 4 in 1/94 (1.1%) and CAD-RADS 5 in 0 subjects.

AI-analysis detected stenosis > 0% in 53/94 (56.4%) subjects with a severity of CAD-RADS 1 in 19/94 (20.2%), CAD-RADS 2 in 22/94 (23.4%), CAD-RADS 3 in 0/94, CAD-RADS 4 in 6/94 (6.4%) and CAD-RADS 5 in 6/94 (6.4%) subjects.

In visual analysis, significant stenosis ≥ 50% in any coronary artery was present in 3/94 (3.2%) subjects. AI-analysis detected significant stenosis ≥ 50% in 7/94 (12.8%).

In visual analysis on vessel level, stenosis > 0% in LAD, CX and RCA was present in 32/94 (34.0%), 10/94 (10.6%) and 7/94 (7.4%), respectively.

AI-analysis detected stenosis > 0% in LAD, CX and RCA in 48/94 (51.1%), 17/94 (18.1%) and 27/94 (28.7%).

In visual analysis, significant stenosis ≥ 50% was present in 3/94 (3.2%) subjects in the LAD, and none in the CX or RCA.

AI-analysis detected significant stenosis ≥ 50% in LAD, CX and RCA in 7/94 (7.4%), 4/94 (4.3%) and 5/94 (5.3%), respectively.

Detailed analysis of AI-based FFR-analysis is provided in Table 3. FFRai < 0.8 was detected with the AI-based model in 11/94 (11.7%) subjects, with 7/94 (7.4%) in LAD, 4/94 (4.3%) in CX, and 5/94 (5.3%) in RCA. Among 12 vessels that were labelled with CAD-RADS 4 or 5 by the AI-model, FFRai was estimated < 0.8 in 11 vessels.

Variables are expressed as raw numbers; numbers in parentheses are percentages.

CAD-RADS: Coronary Artery Disease-Reporting and Data System 2.0; FFR-ai: Artificial intelligence-derived fractional flow reserve.

Details of diagnostic performance are provided in Table 4. On subject level, AI-dependent CAD-RADS evaluation yielded 31 true-positives, 38 true-negatives, 22 false-positives, and 3 false-negatives for the detection of any stenosis > 0%, resulting in a sensitivity, specificity, PPV, NPV, and accuracy of 91.2%, 63.3%, 58.5%, 92.7%, and 73.4%. AI-dependent evaluation for the detection of significant stenosis ≥ 50% yielded 3 true-positives, 82 true-negatives, 9 false-positives, and 0 false-negatives, resulting in a sensitivity, specificity, PPV, NPV, and accuracy of 100%, 90.1%, 25.0%, 100%, and 90.4%.

On vessel level, AI-dependent CAD-RADS evaluation for the detection of any stenosis > 0% in LAD, CX and RCA yielded 29, 8 and 7 true-positives, 43, 75 and 67 true-negatives, 19, 9 and 20 false-positives, and 3, 2 and 0 false-negatives, respectively, resulting in a sensitivity, specificity, PPV, NPV, and accuracy of 90.6%, 69.4%, 60.4%, 93.5%, and 76.6% for LAD, 80.0%, 89.3%, 47.1%, 97.4%, and 88.3% for CX and 100%, 77.0%, 25.9%, 100%, and 78.7% for RCA, respectively.

AI-dependent CAD-RADS evaluation for the detection of significant stenosis ≥ 50% in LAD, CX and RCA yielded 3, 0 and 0 true-positives, 87, 90 and 89 true-negatives, 4, 4 and 5 false-positives, and 0 false-negatives, respectively, resulting in a sensitivity, specificity, PPV, NPV, and accuracy for the LAD of 100%, 95.6%, 42.9%, 100%, and 95.7%. Specificity and NPV for CX and RCA were 95.7% and 100%, and 94.7 and 100%, respectively.

Variables are expressed as percentages; numbers in parentheses are 95% confidence interval.

AI: artificial intelligence; CAD-RADS: Coronary Artery Disease-Reporting and Data System 2.0; CX: circumflex artery; FFR: fractional flow reserve; LAD: left anterior descending artery; NA: not applicable; NPV: negative predictive value; PPV: positive predictive value; RCA: right coronary artery.