Serum metabolomics improves risk stratification for incident heart failure

Rafael R. Oexner,Hyunchan Ahn,Konstantinos Theofilatos,Ravi A. Shah,Robin Schmitt,Philip Chowienczyk,Anna Zoccarato,Ajay M. Shah

DOI: https://doi.org/10.1002/ejhf.3226

2024-04-18

European Journal of Heart Failure

Abstract:This study assessed the predictive value of proton nuclear magnetic resonance [1H‐NMR]‐derived serum metabolomics for the prediction of incident heart failure (HF) within the UK Biobank (UKB) cohort. Risk models were generated based on a well‐validated clinical score (Pooled Cohort Equations to Prevent HF [PCP‐HF]), serum metabolomics, or a combination of both. The serum metabolome was highly associated with incident HF, even after adjustment for clinical HF risk. Comparative analysis of risk models revealed strikingly good performance of simple models based solely on age, sex and the serum metabolome, as depicted in a receiver‐operating characteristic curve or emphasized in excellent HF‐free survival stratification in Kaplan–Meier plots (blue shades correspond to risk quintiles, higher risk: more saturated shades). Aims Prediction and early detection of heart failure (HF) is crucial to mitigate its impact on quality of life, survival, and healthcare expenditure. Here, we explored the predictive value of serum metabolomics (168 metabolites detected by proton nuclear magnetic resonance [1H‐NMR] spectroscopy) for incident HF. Methods and results Leveraging data of 68 311 individuals and >0.8 million person‐years of follow‐up from the UK Biobank cohort, we (i) fitted per‐metabolite Cox proportional hazards models to assess individual metabolite associations, and (ii) trained and validated elastic net models to predict incident HF using the serum metabolome. We benchmarked discriminative performance against a comprehensive, well‐validated clinical risk score (Pooled Cohort Equations to Prevent HF [PCP‐HF]). During a 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). Performance‐optimized risk models effectively retained key predictors representing highly correlated clusters (≈80% feature reduction). Adding metabolomics to PCP‐HF improved predictive performance (Harrel's C: 0.768 vs. 0.755, ΔC = 0.013, [95% confidence interval [CI] 0.004–0.022], continuous net reclassification improvement [NRI]: 0.287 [95% CI 0.200–0.367], relative integrated discrimination improvement [IDI]: 17.47% [95% CI 9.463–27.825]). Models including age, sex and metabolomics performed almost as well as PCP‐HF (Harrel's C: 0.745 vs. 0.755, ΔC = 0.010 [95% CI −0.004 to 0.027], continuous NRI: 0.097 [95% CI −0.025 to 0.217], relative IDI: 13.445% [95% CI −10.608 to 41.454]). Risk and survival stratification was improved by integrating metabolomics. Conclusion Serum metabolomics improves incident HF risk prediction over PCP‐HF. Scores based on age, sex and metabolomics exhibit similar predictive power to clinically‐based models, potentially offering a cost‐effective, standardizable, and scalable single‐domain alternative.

cardiac & cardiovascular systems