Subphenotype heterogeneity to guide predictive enrichment in acute kidney injury: insights from machine learning and target trial emulation

Acute kidney injury (AKI) exhibits substantial heterogeneity in clinical presentation. Previous studies used traditional clustering approaches to identify subphenotypes, often focusing on adverse clinical outcomes, while ignoring that the central goal of subphenotyping is to enable individualized care. In this study, we aimed to identify distinct AKI subphenotypes and evaluate their heterogeneous associations with vasopressor choice and renal replacement therapy (RRT) strategies. This retrospective cohort study used data from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database and included 31,030 patients diagnosed with AKI within 48 h of admission. Generative topographic mapping, a probabilistic unsupervised machine-learning model, was used to identify clusters. We then applied a target trial emulation framework to emulate comparisons of norepinephrine versus vasopressin, RRT modalities, and continuous RRT initiation timing across subphenotypes. Four distinct subphenotypes were identified, with marked differences in clinical features, laboratory abnormalities, and outcomes: a large group with intermediate severity; a hyper-inflammatory subphenotype with marked liver dysfunction and severe AKI; a cardiorenal congestion subphenotype with high cardiovascular comorbidity; and a younger subphenotype enriched for postoperative or trauma patients. Vasopressin was associated with reduced mortality in Subphenotype 2. A more delayed continuous RRT initiation strategy was linked to lower 60-day mortality in Subphenotypes 1 and 2. These associations remained robust in sensitivity analyses. We identified four clinically distinct AKI subphenotypes that demonstrated substantial heterogeneity in their mortality associations with vasopressor use and RRT strategies. These findings could improve prognostication and advance precision medicine in critical care.