Abstract 4113263: Serum Metabolomics Adds Predictive Value to Clinical and Polygenic Risk Scores for Atrial Fibrillation: A UK Biobank study of >240,000 Patients

Background: Atrial fibrillation (AF) poses substantial morbidity and mortality burdens globally. It is critical to identify individuals at high-risk for AF to implement both lifestyle modifications and close monitoring for initiation of oral anticoagulants. Currently, risk prediction in AF relies on clinical scores such as CHARGE-AF. Small molecule metabolites are thought to play a role in AF pathogenesis, but the additional predictive value of metabolomics over clinical and genetic risk factors is unknown. Using the UK Biobank (UKBB), we sought to perform this evaluation. Methods: We included all UKBB participants with complete 1 H-NMR metabolite measurements. We excluded participants with prior history of AF or with incomplete AF Polygenic Risk Score (AF-PRS) or CHARGE-AF components. The final cohort included 240,628 patients. In-patient records, primary care notes, and death registries were queried to identify patients who had developed incident AF within 5 years of enrollment. The cohort was divided into a 80/20:Train/Test split. We compared the performance of an Elastic-Net regularized Cox Proportional Hazards (EN-CPH) model trained on CHARGE-AF, AF-PRS, and the full 170 metabolite panel to a CPH model trained only on CHARGE-AF and AF-PRS. Results: Within 5 years, 4,174 (1.7%) patients developed AF. Compared to the control population, the incident AF cases were more likely to be older (62 vs. 56, p<0.001), male (65% vs 45%, p<0.001), and have higher CHARGE-AF and AF-PRS scores (p<0.001). After training the EN-CPH model on CHARGE-AF, AF-PRS, and 170 metabolites, the final model included 8 metabolites on top of CHARGE-AF (HR: 1.28) and AF-PRS (HR: 1.11). After adjusting for CHARGE-AF and AF-PRS, creatinine level was associated with increased risk of AF (HR: 1.01) while linoleic acid level (HR: 0.985) was associated with decreased risk. Furthermore, total cholesterol, esterified cholesterol, free cholesterol, cholesteryl esters in IDL, omega-6 fatty acids, and cholesterol in large LDL were associated with HR 0.993-0.999. The EN-CPH model out-performed the CPH model without metabolomics on the test set (AUC 0.786 vs 0.753, p < 0.001). Conclusions: The addition of metabolomics to clinical and genomic risk scores improves prediction of 5-year incident AF within a general population. This study highlights the significance of the metabolome as an independent prognosticator of AF risk. Further study of the mechanisms by which selected metabolites alter AF risk is needed.