Abstract
BACKGROUND: Early detection of atrial fibrillation (AF) can prevent AF-related complications. Radiomic analysis of epicardial adipose tissue (EAT) was shown to predict AF recurrence postablation, but only limited data exist regarding left atrial EAT (LA-EAT) radiomic analysis for predicting AF in patients with yet unknown AF. Our aim was to develop prediction model for AF, based on the association of machine learning-based radiomic analysis of LA-EAT and AF. METHODS: Retrospective matched case-control study of patients with and without AF, undergoing noncontrast electrocardiographic (ECG)-gated cardiac computed tomography (CT). Segmentation of LA-EAT and extraction of LA-EAT radiomic features were performed using syngo.via Frontier (Siemens Healthineers, Forchheim, Germany). Univariate analysis identified radiomic features associated with AF. Predictive models for AF were developed via logistic regression and machine learning-based random forest analyses. Models were validated on external cohort of patients with 1:1 AF : control ratio and deployed in a real-world setting with an AF : control ratio of 15:85. RESULTS: The study included 280 patients, 120 with documented AF and 160 matched controls. Based on LA-EAT radiomic features, which were significantly associated with AF, logistic regression and random forest models were constructed and tested on separate internal cohort of patients, yielding area under the curve (AUC) of 0.88 and 0.86, respectively, for prediction of AF. External validation verified these results (AUC 0.84 and 0.78, respectively). Both models were further validated in a real-world setting cohort (AUC 0.85 and 0.81, respectively). CONCLUSIONS: Models, based on LA-EAT radiomic features extracted from noncontrast ECG-gated cardiac CT, could accurately predict AF, suggesting a potential widespread noninvasive method for predicting the presence of AF. CLINICAL REGISTRATION NUMBER: 0281-23-ASF.