435 Discovering novel atrial fibrillation phenogroups: A multi-algorithm approach using real-world claims data

Objectives/Goals: To discover novel phenogroups of atrial fibrillation (AF) from a real-world claims database using a multi-algorithm unsupervised machine learning (ML) approach and to reveal clinical factors associated with each phenogroup using supervised ML. We aim to establish clinically meaningful AF phenogroups to advance precision medicine. Methods/Study Population: We will leverage the Merative MarketScan databases (2016–2024) of real-world claims data with an estimated >600,000 AF patients. Four unsupervised ML clustering techniques will be employed to identify phenogroups: k-means, Gaussian mixture models, agglomerative, and HDBSCAN. Following cluster generation, the algorithms will be evaluated through internal validation via the Davies–Bouldin Index which calculates cluster compactness and separation, and the Adjusted Rand Index to determine stability between algorithms. Then, supervised ML via SuperLearner, which will include advanced algorithms (e.g., gradient boosting machines and support vector machines), will identify clinical factors associated with each phenogroup and will be evaluated using k-fold cross-validation and Shapley values. Results/Anticipated Results: We anticipate that the multi-algorithm approach will yield robust, distinct, and clinically meaningful AF phenogroups. SuperLearner will confirm the key clinical characteristics driving membership in each group (e.g., hypertension and heart failure). These data-driven classifications will provide a robust foundation for subsequent translational analyses. We will perform time-to-event analyses to assess the differential risks of stroke, coronary artery disease and myocardial infarction, heart failure, and rates of healthcare utilization across phenogroups. Next, we will employ the target trial framework to evaluate the comparative effectiveness of anticoagulant therapies on stroke incidence across phenogroups. Discussion/Significance of Impact: This work demonstrates a highly rigorous, multi-algorithm approach to characterize a large, real-world database of AF patients. By constructing novel AF phenogroups, we aim to move beyond the limitations of current AF classifications and advance precision medicine by providing the foundation for individualized therapeutic strategies.