Conyers, Frank

Introduction: Embolic stroke of undetermined source (ESUS) is a clinical construct encompassing heterogenous possible mechanisms of incident stroke, potentially with variable responsiveness to anticoagulation and antiplatelet treatment. This study aims to identify differences in recurrent stroke risk based on phenotypic clusters in an ESUS population, and whether some clusters may benefit relatively more from anticoagulation over antiplatelet therapy. Methods: We identified patient phenotypes from a multicenter, retrospective cohort study of 1909 ESUS patients by dividing them into groups using Partitioning Around Medoids (PAM) with correlation-weighted Gower distance over 23 baseline demographic, clinical, laboratory, radiographic, and cardiac features. We confirmed optimal clustering at 3 groups. We assessed cluster associations with time to recurrent stroke using Kaplan-Meier survival analysis. We then employed a causal survival forest to estimate conditional average treatment effects (CATE) of anticoagulation versus antiplatelet therapy using the same baseline features as predictors with a 365-day horizon for restricted mean survival time differences. Finally, we used best linear projection to identify baseline features predictive of treatment effect and assessed whether effects differed significantly by cluster. Results: Cluster 1 (n=542) identified a low-risk, younger sample with minimal baseline disability and comorbidity burden. Cluster 2 (n=743) comprised a moderate-risk phenotype with intermediate comorbidity. Cluster 3 (n=626) represented the oldest and highest-risk phenotype. Kaplan-Meier analysis demonstrated significant differences in stroke-free survival across patient clusters (p < 0.001), where Cluster 1 experienced the lowest recurrent stroke rate and Cluster 3 had the highest. The clustering algorithm successfully identified clinically meaningful phenotypes with distinct risk profiles. While CATE varied across clusters, the differences were not statistically significant. Conclusion: In a multicenter ESUS population, we identified 3 distinct phenotypic clusters, each comprising nearly one-third of the sample and with a differential risk of recurrent stroke. Although cluster was not a significant effect modifier, substantial CATE variability by cluster was observed. Future clinical trials may consider evaluating secondary stroke prevention therapies among phenotypically similar patients who may respond differentially to novel therapies.