Machine-learning-derived radiomics signature of pericoronary tissue in coronary CT angiography associates with functional ischemia

Abstract

Objectives: To determine the association between radiomics signature (Rad-signature) of pericoronary tissue (PCT) in coronary computed tomography angiography (CCTA) and CT-derived fractional flow reserve (CT-FFR), and explore the influential factors of functional ischemia. Methods: We retrospectively included 350 patients who underwent CCTA from 2 centers, consisting of the training (n = 134), validation (n = 66), and testing (with CCTA and invasive coronary angiography, n = 150) groups. After evaluating coronary stenosis level in CCTA (anatomical CT), pericoronary fat attenuation index (FAI), and CT-FFR, we extracted 1,691 radiomic features from PCT. By accumulating and weighting the most contributive features to functional ischemia (CT-FFR ≤ 0.8) the Rad-signature was established using Boruta integrating with a random forest algorithm. Another 45 patients who underwent CCTA and invasive FFR were included to assure the performance of Rad-signature. Results: A total of 1046 vessels in 350 patients were analyzed, and functional ischemia was identified in 241/1046 (23.0%) vessels and 179/350 (51.1%) patients. From the 47 features highly relevant to functional ischemia, the top-8 contributive features were selected to establish Rad-signature. At the vessel level, the area under the curve (AUC) of Rad-signature to discriminate functional ischemia was 0.83, 0.82, and 0.82 in the training, validation, and testing groups, higher than 0.55, 0.55, and 0.52 of FAI (p < 0.001), respectively, and was higher than 0.72 of anatomical CT in the testing group (p = 0.017). The AUC of the combined model (Rad-signature + anatomical CT) was 0.86, 0.85, and 0.83, respectively, significantly higher than that of anatomical CT and FAI (p < 0.05). In the CCTA-invasive FFR group, using invasive FFR as the standard, the mean AUC of Rad-signature was 0.83 ± 0.02. At the patient level, multivariate logistic regression analysis showed that Rad-signature of left anterior descending (LAD) [odds ratio (OR) = 1.72; p = 0.012] and anatomical CT (OR = 3.53; p < 0.001) were independent influential factors of functional ischemia (p < 0.05). In the subgroup of nonobstructive (stenosis <50% in invasive coronary angiography) and obstructive (≥50%) cases of the testing group, the independent factor of functional ischemia was FAI of LAD (OR = 1.10; p = 0.041) and Rad-signature of LAD (OR = 2.45; p = 0.042), respectively. Conclusion: The machine-learning-derived Rad-signature of PCT in CCTA demonstrates significant association with functional ischemia.

Keywords: coronary computed tomography angiography; fractional flow reserve; functional ischemia; machine learning; radiomics.