Diagnosing atrial fibrillation early: the power of radiomics analysis on non-contrast cardiac CT

Abstract

Abstract Background Often silent and asymptomatic, atrial fibrillation (AF) contributes significantly to morbidity and mortality. Early detection of AF could prevent and treat AF complications. Epicardial adipose tissue (EAT) is a unique fat tissue located on the surface of the myocardium that can be demonstrated on gated non-contrast cardiac CT (CCT). Radiomics, a method that translates medical images into quantitative data, can yield biological information and enable radiologic phenotypic profiling. Purpose This study aimed to identify radiomic features of both EAT and left atrium (LA)-EAT in patients with AF, based on non-contrast CCT, and utilize these features to construct a prediction model for the presence of AF. Methods We conducted a retrospective case control study of 180 patients who underwent cardiac CTA (CCTA) for clinical indications between 2018-2022 in Shaare Zedek Medical Center Jerusalem. The AF group contained 90 patients with known AF referred to CCTA prior to pulmonary vein isolation. The control group included 90 patients referred to CCTA for other indications, without a definite or suspected AF diagnosis. In all patients, the cardiac CT protocol included initial non-contrast CCT, which we used for radiomic analysis. The study dataset was trained on 60 patients from each group, with evaluation performed on a validation dataset of 30 patients each. Detection of the cardiac EAT was performed fully automatically using the cardiac risk analysis tool, SyngoVia, Research Frontier, Siemens Healthineers. Additional specific segmentation of the LA- EAT was performed semi-automatically using fixed anatomical landmarks. Radiomic analysis was performed via Radiomic tool, SyngoVia, Research Frontier, Siemens Healthineers. Results Both groups included in this research, study and control, had matching clinical characteristics such as age (average 63.2) , sex (female 50%), history of dyslipidemia, hypertension, diabetes, and smoking. A univariant statistical analysis for detection of AF of 1691 radiomic features in LA-EAT found an AUC of 0.866, p<<0.001 and 0.804, p<<0.001 for morphological and texture features, respectively. In a multivariate regression analysis, combining the top two radiomic features increased the AUC to 0.89, p = 0.004. Evaluation of linear regression and random forest model applications on a new validation cohort of 60 patients verified the results with AUC curves of 0.88 and 0.87 respectively. Compared to LA-EAT, applying total EAT prediction model on a new validation cohort produced inferior results, with AUC curves of 0.79 and 0.8 for linear regression and random forest model applications, respectively. Conclusion This proof-of-concept study demonstrated that radiomic features of LA-EAT are associated with and predictive of AF. Models based on radiomic features from gated non-contrast CCT of LA-EAT identified the presence of AF with excellent diagnostic performance. Models based on total EAT produced inferior results.Figure 1Heatmap of radiomic featuresFigure 2ROC curves