Utilization of Serum Metabolomics and Polygenic Risk Scores in a Novel Risk Stratification Tool for the Prediction of Incident Atrial Fibrillation

BACKGROUND: Atrial fibrillation (AF) is associated with substantial morbidity and mortality. We sought to investigate the predictive value of serum metabolomics for 5-year incident AF in the context of clinical and polygenic risk score (PRS) stratification tools. METHODS: We studied a cohort of 240 628 patients UK Biobank participants with proton nuclear magnetic resonance spectroscopy measurements of 170 serum metabolites at enrollment. Five-year incidence of AF was assessed using Cox proportional hazards models. Cohorts for Heart and Aging Research in Genomic Epidemiology-AF and AF polygenic risk score (AF-PRS) scores were used as benchmark risk models for comparison. Models were trained on 80% of the cohort, and performances were validated on the remaining 20% cohort. Performance of clinical, AF-PRS, and combined metabolomics models was evaluated using time-dependent area under the receiver operating characteristic curve, net reclassification improvement, and relative integrated discrimination improvement analysis. RESULTS: During follow-up, 4174 (1.7%) participants developed AF. After training a model on the full metabolomics panel in addition to Cohorts for Heart and Aging Research in Genomic Epidemiology-AF and AF-PRS, the final model retained 8 metabolites. Creatinine level was associated with increased risk (hazard ratio, 1.01 per 1 SD log-transformed value [95% CI, 1.00–1.03]) while linoleic acid level (hazard ratio, 0.985 [0.979–0.994]) was associated with decreased risk of AF. The addition of metabolomics to the Cohorts for Heart and Aging Research in Genomic Epidemiology-AF+AF-PRS model improved risk prediction (5-year time-dependent area under the receiver operating characteristic curve, 0.789 [0.776–0.802] versus 0.755 [0.738–0.772]; P <0.05) and stratification on the validation set (NRI cases : 11.1%, NRI controls : 3.1%, IDI relative : 11.6%). A model using only age, sex, metabolomics, and AF-PRS had fair risk prediction on the validation set (5-year time-dependent area under the receiver operating characteristic curve, 0.787 [0.773–0.801]). CONCLUSIONS: The addition of metabolomics to clinical and genomic risk scores improves the prediction of 5-year incident AF. A risk stratification tool using age, sex, and serum metabolomics and AF-PRS provides excellent AF risk prediction. Mechanisms by which specific metabolites reflect AF risk require further exploration.