Deep Learning Assessment of Epicardial Adipose Tissue: New Perspectives on Major Adverse Cardiac Events Prediction

Epicardial Adipose Tissue (EAT) is a visceral fat depot surrounding the heart, playing a crucial role in cardiac pathophysiology mechanisms. Its local effect also contributes to cardiac conditions such as coronary artery disease, atrial fibrillation, and heart failure. The relationship between EAT characteristics and major adverse cardiac events (MACE) has emerged as a significant area of research. However, the underlying predictive pattern still remains to be understood. The aim of this study is to develop an autoencoder (AE) architecture to reconstruct cardiac magnetic resonance images (CMR) of EAT and extract its most representative features to classify between patients with MACE event and event-free subjects. An AE architecture was built to get a compressed image representation of 251 patients' images. Ninety patients, 45 experiencing a MACE event and 45 event-free, were included in the subsequent analysis. A total of 128 features were extracted to train and test different machine learning models. A stratified 80-20% training-test split was performed, with additional validation partition from the training set. Five features selection methods (Mann Whitney U test, least absolute shrinkage and selection operator (LASSO), semi-supervised LASSO, principal component analysis (PCA), and semi-supervised PCA) and three classifiers (support vector machine, logistic regression, gradient boosting) were combined and evaluated on the validation set. Based on this assessment, supervised PCA and logistic regression were chosen to build the final model which reached on the test set an accuracy, sensitivity and specificity of 0.89, 1 and 0.78, respectively. These preliminary results showed the potential of this deep learning approach to predict the occurrence of MACE event starting from EAT images.Clinical Relevance - This work establishes the efficacy of deep learning-based features obtained from epicardial adipose tissue images to differentiate between patients with and without MACE event.