Machine-learning facilitates selection of a novel diagnostic panel of metabolites for the detection of heart failure.

The metabolic derangement is common in heart failure with reduced ejection fraction (HFrEF). The aim of the study was to check feasibility of the combined approach of untargeted metabolomics and machine learning to create a simple and potentially clinically useful diagnostic panel for HFrEF. The study included 67 chronic HFrEF patients (left ventricular ejection fraction-LVEF 24.3 ± 5.9%) and 39 controls without the disease. Fasting serum samples were fingerprinted by liquid chromatography-mass spectrometry. Feature selection based on random-forest models fitted to resampled data and followed by linear modelling, resulted in selection of eight metabolites (uric acid, two isomers of LPC 18:2, LPC 20:1, deoxycholic acid, docosahexaenoic acid and one unknown metabolite), demonstrating their predictive value in HFrEF. The accuracy of a model based on metabolites panel was comparable to BNP (0.85 vs 0.82), as verified on the test set. Selected metabolites correlated with clinical, echocardiographic and functional parameters. The combination of two innovative tools (metabolomics and machine-learning methods), both unrestrained by the gaps in the current knowledge, enables identification of a novel diagnostic panel. Its diagnostic value seems to be comparable to BNP. Large scale, multi-center studies using validated targeted methods are crucial to confirm clinical utility of proposed markers.