Improving cardiovascular risk assessment through comprehensive radiomics analysis of epicardial adipose tissue in screening non-contrast CT calcium score images

Cardiovascular disease is the major cause of death, with projected increases due to the aging population and obesity. Agatston scoring of CT calcium score (CTCS) images is widely used to predict the risk of a future major adverse cardiovascular event (MACE) and provides therapeutic guidelines. Our team opportunistically analyzed epicardial adipose tissue (EAT) in CTCS images to improve MACE risk prediction. Here, we expand EAT analysis by incorporating radiomics textural features. We used 2,293 CTCS images, with 13.65% MACE from the UH-CLARIFY registry, a no-cost screening program at University Hospitals of Cleveland. The dataset was divided into balanced training and testing sets (70%/30%). EAT volumes were automatically segmented using our previously reported DeepFAT software. Textural features were derived by applying 160 radiomic filters (e.g., Haralick, Gabor, Laws, and CoLlaGe, at various window sizes) to capture patterns of intensity changes that potentially characterize the inflammatory profile of EAT. Statistical measurements—median, variance, kurtosis, and skewness—were used to aggregate response volumes producing a total vector of 640 textural features. For our feature selection, we employed mRMR feature preselection and elastic-net Cox to identify a prognostic feature set for time-to-event modeling. Using Cox time-to-event modeling, we compared the predictive power of Agatston, EAT volume, and textural features, individually and in combination with clinical data. We identified eight texture features that were strongly prognostic of MACE. The comprehensive model incorporating texture achieved superior performance on the held-out test data, with AUC/C-index of 73.2/73.6, outperforming Agatston (70.6/70.5) and EAT volume (56.1/56.7) despite the well-known resistance-to-change of C-index in low-event rate data. Further, the Akaike Information Criterion (AIC) supported the inclusion of textures, which enhances the predictive power of the model. This preliminary study shows that textures in EAT, possibly responding to inflammation, are predictive of MACE, suggesting their use in comprehensive cardiometabolic risk prediction for personalized medicine.