A17-06 Longitudinal, Outcome-Aware Acute Hypoxic Respiratory Failure Phenotypes: Discovery, External Validation, and Early Prediction

Abstract Rationale Acute Hypoxaemic Respiratory Failure (AHRF) is biologically heterogeneous, yet most phenotyping is cross-sectional and limited to the first few ICU days. We sought longitudinal, outcome-aware trajectory phenotypes of persistent AHRF, and to forecast class membership early from routine data. Methods We analysed adults with persistent AHRF (PaO2/FiO2 P/F 300 mmHg and PEEP ≥5 cmH2O for ≥72 h) from three ICU databases: MIMIC-IV (USA; derivation, n = 3,938), Imperial College Healthcare NHS Trust (UK; validation, n = 2,888), and AmsterdamUMCdb (Netherlands; validation, n = 3,592). Daily mean P/F to day 14 and times to ICU discharge or death were jointly modelled using a competing-risk latent-class mixed model. The optimum number of classes was chosen by Bayesian information criterion/entropy. The fitted model was frozen and applied unchanged to the validation cohorts. Prediction of class membership as early as possible from onset of AHRF using a 12-variable XGBoost model. Generalisability was also examined in an expert-annotated ARDS subset. An open-source clinical classifier provided exploratory hyper-/hypoinflammatory phenotyping. Results A four trajectory class (TC) solution provided optimal fit (high entropy; mean posterior probability 0.80) and clinically coherent trajectories with distinct outcomes (Figure 1). TC1: Early-recovery (n = 764): 100% discharged by day 14. TC2: Stable persistence (n = 1,668): flat P/F; 8% mortality and 33% discharge by day 14. TC3: Biphasic improvement-deterioration (n = 754): fastest early P/F rise, then decline; 17% mortality and 57% discharge by day 14. TC4: Rapid-decline (n = 752): near-monotonic P/F fall; 96.5% mortality by day 14. Although learned from P/F and outcomes alone, classes differed in other organ-dysfunction markers (e.g., vasopressor use, lactate, renal dysfunction). The frozen model generalised with high assignment certainty in ICHT (mean posterior 0.87) and AmsterdamUMCdb (0.81), with close agreement between predicted and observed P/F and outcome curves. Early prediction achieved AUCs ≥0.77 using data up to day 3. ARDS patients were distributed across all classes and were enriched in TC2 (50% vs 33% in non-ARDS). At onset of AHRF, hyperinflammatory prevalence was 32% in MIMIC and 28% in ICHT and was enriched in TC4 (46% MIMIC; 53% ICHT) and lowest in TC1 (19% MIMIC; 18% ICHT). Conclusions To our knowledge, these are the first reproducible AHRF P/F trajectories that extend beyond the initial few ICU days while accounting for competing risks. They generalise across three cohorts, are predictable within 72h using routine data, and span both ARDS status and inflammatory subphenotypes. Trajectory phenotyping offers a scalable framework for early risk stratification, biologically integrative studies, and trial enrichment. This abstract is funded by: Medical Research Council UK