Enhanced Early Risk Stratification in Cardiogenic Shock Using Machine Learning: A Hierarchical Clustering Analysis

Cardiogenic shock (CS) exhibits marked clinical heterogeneity, limiting the performance of conventional risk stratification tools like the Society for Cardiovascular Angiography and Interventions (SCAI) staging system. Machine learning (ML) has identified reproducible CS phenotypes that improve risk prediction, but their temporal consistency across the progression of shock remains incompletely described.

The purpose of this study was to evaluate the temporal evolution of ML-derived CS phenotypes across the first 3 intensive care unit (ICU) days and determine whether phenotype trajectories enhance mortality stratification beyond SCAI staging.

In a multicenter retrospective registry of 2,906 adults with CS admitted to ICUs between 2016 and 2022, hierarchical agglomerative clustering was applied independently to clinical variables on ICU days 1, 2, and 3.

On day 1, 3 phenotypes emerged: Cardiac (younger, acute myocardial infarction-predominant, preserved renal function; 15% mortality), Cardio-Renal (older, non-acute myocardial infarction-predominant, isolated renal extracardiac impairment, predominantly SCAI A and B on admission; 27% mortality), and Cardio-Metabolic (severe renal and tissue malperfusion, predominantly SCAI D and E on admission; 55% mortality). Similar phenotypes persisted on day 2 and day 3. These ML-derived phenotypes enhanced mortality stratification within each SCAI stage across days 1 to 3.

Hierarchical clustering reproduced previously identified CS phenotypes that were consistent across the first 3 days of ICU stay, providing an enhanced framework for early risk stratification. Prospective validation is required for clinical implementation.