Abstract Body (Do not enter title and authors here): Introduction: Accurate risk prediction is needed to identify and prevent heart failure events and disease progression. Recently, we completed the metabolomic profiling of more than half a million UK Biobank blood samples, doubling previously available metabolite data.

Aims: Here, we compared the use of a metabolomic score, the clinical Framingham risk score and a polygenic score in predicting incident heart failure and assessed portability by stratifying performances by genetic ancestry. Using only metabolomic and genomic scores, we predicted progression to death within individuals with a prior diagnosis.

Methods: We used baseline metabolite data from the UK Biobank measured with nuclear magnetic resonance spectroscopy, which include a range of routine lipids, fatty acids and amino acids. We derived the polygenic score using PRS-CS on genome-wide association study results from the Global Biobank Meta-analysis Initiative. Individuals with prevalent heart failure or other cardiovascular disease were excluded from the incidence analysis. In half the data we trained models using Cox regressions and LASSO regularisation with 10-fold cross-validation. In the remaining data, we tested combined risk scores using age- and sex-adjusted hazard ratios and AUCs.

Results: Within 10 years of follow-up, there were 2,855 incident heart failure cases. Framingham (AUC 0.745) performed better than the polygenic risk score (AUC 0.736), but not as well as the metabolomic score (AUC 0.769) in European ancestry individuals. The same pattern was observed for individuals with South-Asian or African ancestry, though clinical and metabolomic predictions were very similar for the latter. Adding metabolomics on top of Framingham resulted in delta AUCs of 0.034, 0.031 and 0.021 for European, South-Asian and African ancestries, respectively. There were 4,236 individuals with a heart failure diagnosis prior to blood sampling, of which 122 progressed to death. Disease progression was not predicted by the polygenic score, but using metabolomics, hazard ratios were 1.56 [95% CI 1.30-1.87] per SD and 3.57 [2.74-5.30] for the top decile versus remaining 90%.

Conclusions: In summary, the use of metabolomic scores can not only improve heart failure incidence and progression predictions, but also has the strongest standalone prediction across genetic ancestries. Additionally, enhancing clinical risk prediction with metabolites could make screening efforts more effective and extend portability.