Abstract Body (Do not enter title and authors here): Background: Atrial myopathy—characterized by abnormal left atrial (LA) function or size— is independently associated with cardiovascular outcomes such as atrial fibrillation (AF) and stroke. One obstacle to the discovery of strategies to prevent atrial myopathy-related outcomes is its heterogeneity.

Objective: Resolve heterogeneity of atrial myopathy by identifying distinct subtypes of atrial myopathy with differential risk of AF using an unsupervised machine learning clustering analysis.

Methods: We included ARIC participants at visit 5 (2011-13) with atrial myopathy—defined using LA reservoir strain (Female: <28.6%, Male: <26.4%)—and no prevalent AF. We considered 52 clinical, demographic, ECG, and echocardiographic candidate variables for the clustering analysis. Random forest imputation was used for missing data. LASSO regression of atrial myopathy was used to select pertinent variables. Selected variables were standardized, and clusters were identified using model-based clustering with mClust, based on the Bayesian Information Criterion. Multivariable Cox regression model was used to evaluate the association between final clusters and incident AF.

Results: In total, 1,415 (mean age ± SD, 77 ± 5 years, 70% Female, 26% Black) participants with atrial myopathy were included in the clustering analysis, and 5,209 participants in the Cox regression analysis. There were 22 variables selected, and 3 clusters identified (Table 1). Over a mean follow-up of 6.3 years, 712 incident AF cases occurred. Compared to those with no atrial myopathy, all three clusters had differentially increased risk of incident AF (Table 2).

Conclusion: Clustering analysis of individuals with atrial myopathy identified 3 unique clusters that were differentially associated with incident AF providing evidence of the heterogeneity of atrial myopathy. The clustering scheme may be used to select high-risk patients for clinical trials aimed at preventing outcomes related to atrial myopathy.