LAUSR

Introduction: Atrial Fibrillation (AF) is a heterogeneous arrythmia. Identification of subgroups of patients with shared mechanisms, such as genetic predisposition to AF and inflammation, could facilitate precision care. Machine learning (ML) is increasingly applied to define disease subtypes in an unbiased fashion, but the extent to which purported subgroups are robust across ML methods and whether they align with underlying AF mechanisms is unclear. Objective: Compare AF subtypes identified using 2 ML methods, and assess the extent to which subtypes relate to genetically-predicted AF and inflammatory state. Methods: We identified individuals with AF in a large electronic health record and linked DNA biobank. Using 86 curated clinical features, 5 clusters were found using affinity propagation, which provided K for a complementary K-means clustering. Shared cluster membership was assessed with the Adjusted Rand Index (ARI). Polygenic risk scores (PRS) for AF and 5 inflammatory biomarkers (IL-1b, IL-6, TNF-α, IFN-γ, IL-10, and IL-17) were used to interrogate genetic mechanisms underlying clustering. Results: We identified 49,905 patients with AF, including 5,532 with genotyping. There was moderate agreement in subgroup membership between the two ML methods (ARI 0.53). Subgroups were distinguished by presence/absence of structural heart disease, comorbidities, mortality, AF ablation, bleeding events, and dialysis (Figure 1). All subgroups had greater mean AF PRS than non-AF controls. Among AF subgroups, the “high comorbidities” group had the lowest AF PRS values (Figure 2). PRS of inflammatory biomarkers were similar between clusters. Conclusion: In this real-world AF cohort, there was moderate agreement between 2 clustering methods. The genetically-predicted levels of inflammatory biomarkers were not different among clusters. Clusters have features comparable to clinically recognized AF subgroups, and differ in genetic liability toward AF.