Abstract Rationale Acute respiratory distress syndrome (ARDS) remains heterogeneous, and trial-derived latent phenotypes are rarely reproduced in electronic health record (EHR) cohorts. We aimed to recover prognostically meaningful ARDS phenotypes from MIMIC-IV and test whether latent profile analysis (LPA) as applied to routinely collected EHR data could identify reproducible ARDS subphenotypes in MIMICIV, and whether these phenotypes correlated with clinical outcomes. Methods We performed a retrospective cohort study using MIMICIV v3.1. ARDS was defined with a computable approach utilizing Berlin and Global ARDS oxygenation criteria. We fit a twoclass LPA using 25 prespecified indicators spanning demographics, vital signs, laboratory values, ventilatory parameters, and key diagnoses/exposures. Outcomes included 90day and in-hospital mortality, invasive organ-support use, and subgroup descriptors (e.g., social determinants of health). Results Among 24,363 ARDS admissions, 9,268 (38.1%) were hyperinflammatory and 15,095 (61.9%) hypoinflammatory. Hyperinflammatory admissions were younger and included significantly more women (40.5% vs 36.0%, p 0.001) and Black patients (10.9% vs 7.5%, p 0.001), as well as worse oxygenation and greater physiologic derangement at baseline. Survival differed by phenotype (log-rank p 0.001): 90-day mortality was 40.8% in the hyperinflammatory class versus 19.1% in the hypoinflammatory class (adjusted OR, 2.19; 95% CI, 2.05-2.35), and restricted mean survival time over 90 days was 11.9 days shorter in the hyperinflammatory group. In-hospital mortality remained higher for hyperinflammatory admissions (57.0% vs 33.3%, p 0.001), reflecting stays beyond 90 days. In adjusted analyses, the hyperinflammatory phenotype was independently associated with greater need for advanced organ support, including invasive mechanical ventilation (80.0% vs 37.5%; adjusted OR, 5.67; 95% CI, 5.32-6.04), tracheostomy, vasopressor support, and renal replacement therapy. Conclusions This study shows that two clinically distinct ARDS subphenotypes can be reproduced at scale in an unselected EHR population using a latent profiling approach. The hyperinflammatory state demonstrated higher acuity at presentation and was consistently linked to materially worse outcomes, including an approximately twofold increase in 90-day mortality (21.7 pp absolute delta), shorter restricted mean survival over 0-90 days, and greater use of advanced organ support. These findings support phenotype-guided risk stratification and provide a basis for integrating subphenotypes into decision-support workflows to trigger targeting or enrollment in prospective, phenotype-directed trials. Differences in sex and race/ethnicity further underscore the need for equity safeguards as phenotype-based care pathways are developed. This abstract is funded by: None
Krishna et al. (2026) studied this question.