Serum metabolomics improve risk stratification for incident heart failure

Abstract

AbstractBackground and AimsPrediction and early detection of heart failure (HF) is crucial to mitigate its impact on quality of life, survival, and healthcare expenditure. In this study, we explored the predictive value of serum metabolomics (168 metabolites detected by proton nuclear magnetic resonance (1H-NMR) spectroscopy) for incident HF.MethodsWe leveraged data of 68,311 individuals and > 0.8 million person-years of follow-up from the UK Biobank (UKB) cohort to assess individual metabolite associations and to train models to predict HF risk in individuals not previously considered at risk. Specifically, we (I) fitted per-metabolite COX proportional hazards (COX-PH) models to assess individual metabolite associations and (II) trained and internally validated elastic net (EN) models to predict incident HF using the serum metabolome. We benchmarked discriminative capacities against a comprehensive, well-validated clinical risk score (Pooled Cohort Equations to Prevent HF, PCP-HF1).ResultsDuring median follow-up of ≈ 12.3 years, several metabolites showed independent association with incident HF (90/168 adjusting for age and sex, 48/168 adjusting for PCP-HF; false discovery rate (FDR)-controlled P < 0.01). Performance-optimized risk models effectively retained key predictors representing highly correlated clusters (≈ 80 % feature reduction). The addition of metabolomics to PCP-HF improved predictive performance (Harrel’s C: 0.768 vs. 0.755.; continuous net reclassification improvement (NRI) = 0.287; relative integrated discrimination improvement (IDI): 17.47 %). Simplified models including age, sex and metabolomics performed almost as well as PCP-HF (Harrel’s C: 0.745 vs. 0.755, continuous NRI: 0.097, relative IDI: 13.445 %). Risk and survival stratification was improved by the integration of metabolomics.ConclusionsThe assessment of serum metabolomics improves incident HF risk prediction. Scores based simply on age, sex and metabolomics exhibit similar predictive power to clinically-based models, potentially offering a cost- and time-effective, standardizable, and scalable single-domain alternative to more complex clinical scores.