Radiomic analysis of Epicardial Adipose Tissue in cardiac MRI for hospitalization risk assessment

Epicardial Adipose Tissue (EAT) is a visceral fat depot whose local effects on cardiac function are incorporated into the complex pathomechanism of coronary artery disease, atrial fibrillation, and heart failure. The aim of this study is to create a radiomic signature to assess the hospitalization risk of 261 patients, from EAT cardiac magnetic resonance-based radiomics. 107 radiomic features were extracted from EAT segmented images and their robustness was assessed by means of ROI translations and intraclass correlation coefficient computation. Robust features were subjected to further selection steps comprising a statistical test and a univariate Cox regression model. Selected features were used to create the radiomic signature, testing three survival models: Cox Proportional Hazard (CPH) regression, random survival forest (RSF), and deep survival analysis (DeepSurv). Leave-one-out cross-validation was performed to build Kaplan-Meier curves and evaluate models' performances. Ten clinical variables were then added to the analysis to test that radiomic signature maintained its significative prognostic value. 56 radiomic features were identified as robust and two textural features, Gray Level Non-Uniformity Normalized and Cluster Shade, were kept to create the radiomic signature. CPH and DeepSurv significantly identified patients with a high risk of hospitalization. Fitting the CPH model with radiomic and clinical features, radiomic signature coefficients remained significant (p=0.04). In conclusion, results achieved in this study have shown that EAT-based radiomics could be a promising tool in assessing patient hospitalization risk.