Machine learning profiling of epicardial adipose tissue using multi-omics data from ventricular biopsies

Epicardial adipose tissue (EAT) has been linked with ventricular arrhythmias, coronary artery disease (CAD) and heart failure (HF). The underlying mechanisms remain poorly understood. Also, predicting the EAT profile based on biopsies could serve as a proxy for the risk of developing heart disease.

To build a machine learning model for predicting progressive increase in EAT ventricular adiposity, based on differentially expressed transcriptomic, metabolomic and lipidomic elements.

Human donor’s hearts (N=59) were classified based on the EAT percentage coverage on the anterior ventricular surface. Tissue samples from the basal anterior left ventricular region provided transcriptomic (bulk RNA sequencing), metabolomic, and lipidomic data. DESeq2, limma and lipidr library algorithms revealed genes, metabolites and lipids differentially expressed per unit of epicardial adipose tissue increase. Training and testing of five regressors were evaluated by their mean absolute error (MAE) and their R2. These were random forest (RF), linear regressor, support vector machines, XGBoost and AdaBoost. Grid search serves for hyperparameter tuning. Each model underwent a 10-fold cross-validation. The omics elements were features and the EAT percentage was the output.

The EAT percentages varied between 70% and 100% with a standard deviation of 12.6 and a mean value of 79.8. The median value was 82. There were 34 significantly expressed genes (Table). Among them were genes related to CAD (CA3, PCSK1 and SPP1), inflammation (CHIA and IL11) and electrical instability (KCNK3). No metabolites or lipids were significantly expressed based on the adjusted p-values. 6 metabolites (Table) and 154 lipids had nominally significant p values (<0.05). Although the metabolite analysis could have identified these six associations by chance, it is important to note that phosphoenolpyruvate (PEP) is linked to heart failure. The switch from fatty acid oxidation to a glycolytic pathway is seen in HF. The PEP from the analysis highlights the link between EAT increase and HF. The most performant models on the multi-omics data were RF and AdaBoost (Figure). Their MAE was lower than the standard deviation of the data set. Conclusion: The hearts with increased EAT percentage expressed genes and metabolites linked to CAD, arrhythmias and HF. Multi-omics data from ventricular biopsies can be used for EAT profiling. This might serve as a proxy for the risk of developing heart disease.