From Clusters to Outcomes: Machine Learning-Based Phenotyping in Intermediate-High-Risk Acute Pulmonary Embolism.

Intermediate-high-risk (IHR) pulmonary embolism (PE) represents a heterogeneous group in whom guideline-based criteria may insufficiently capture biologic and hemodynamic variability relevant to early deterioration. Data-driven phenotyping may improve risk stratification and support individualized decisions regarding reperfusion therapy. In this retrospective cohort study (2012-2025), 553 guideline-defined IHR PE patients were analyzed using unsupervised machine learning. Thirty-six demographic, clinical, laboratory, echocardiographic, and CT variables were standardized and encoded as appropriate for clustering. Multiple algorithms were compared, and the optimal model was selected using silhouette width and stability metrics. Clinical characteristics, imaging findings, treatment patterns, and outcomes were compared across phenotypes. The primary outcome was in-hospital mortality; secondary outcome was all-cause long-term mortality. Multivariable logistic regression and Cox models assessed associations with outcomes, and pre-post-treatment changes were evaluated. Two phenotypes were identified using the k-prototypes algorithm (silhouette width = 0.697). Cluster 1 (RV-failure phenotype;