Abstract
ABSTRACT
Objective
To evaluate treatment effect heterogeneity across phenotypic clusters of patients with cryptogenic embolic stroke using causal inference methods.
Methods
From a retrospective, multicenter cohort of patients with cryptogenic embolic stroke, we estimated the treatment effect of anticoagulation versus antiplatelet therapy in patients with ESUS in two ways: (1) Partitioning Around Medoids and (2) causal survival forest modeling. We assessed cluster‐level associations with time to primary outcome (recurrent stroke, major bleeding, or death) using Kaplan–Meier survival analysis to assess cluster differences. We then extracted SHapley Additive exPlanations (SHAP) feature importances from the causal forest model to identify candidate effect modifiers from the causal forest.
Results
Among 1869 patients with ESUS, three phenotypically distinct clusters were extracted with significant differences in event‐free survival across clusters (log‐rank
p
p
= 0.23). SHAP feature importance analysis identified several candidate effect modifiers; however, only higher LVEF (
β
= −327, 95% CI: −580 to −74,
p
= 0.011) and cancer history (
β
= −226, 95% CI: −441 to −11,
p
= 0.039) were independent predictors of response to antiplatelet over anticoagulation using best linear projections. Phenotypic cluster assignment did not significantly predict treatment effect heterogeneity.
Interpretation
Three distinct phenotypic clusters of cryptogenic stroke patients emerged with a differential risk of the composite outcome, but they did not predict responsiveness to anticoagulation. LVEF and cancer history, however, showed differential outcomes across antithrombotic strategies. These findings demonstrate the potential of causal inference methods for identifying effect heterogeneity and individualizing treatment strategies.