Investigation of the relationship between structural features of epicardial adipose tissue and coronary artery disease using machine learning methods

Aim: The aim of this study was to segment epicardial adipose tissue (EAT) from non-contrast computed tomography (CT) images obtained in the coronary artery calcium scoring protocol, to extract its structural features, and to evaluate the value of these features—together with total calcium score (TCS) and cardiovascular risk factors—in predicting the severity of coronary artery disease (CAD) using machine learning methods. Materials and Methods: Between January 2021 and May 2025, coronary computed tomography angiography (CTA) examinations of 417 patients with suspected CAD were retrospectively analyzed. Non-contrast CT images from the calcium scoring protocol were used, and EAT was manually segmented. Structural features extracted from the segmentation data were combined with TCS and cardiovascular risk factors for analysis. Using these data, both deep learning algorithms (unimodal and multimodal models) and classical machine learning methods were developed to predict CAD severity. The diagnostic performance of the models was evaluated using standard metrics, including sensitivity, specificity, accuracy, F1-score, and the area under the receiver operating characteristic curve (AUC). Results: Deep learning models based on the U-Net3D architecture demonstrated high performance in predicting cases with CAD-RADS ≥ 3, which was defined as the primary endpoint of the study. The best performance was achieved in the multimodal model, where the integration of segmentation data with cardiovascular risk factors and TCS significantly improved model accuracy. Conclusion: This study demonstrated that artificial intelligence–based models integrating structural features of EAT with total calcium score and fundamental cardiovascular risk factors can effectively predict the severity of coronary artery disease.