Early PaO₂/FiO₂ trajectory stratifies mortality risk in acute hypoxemic respiratory failure: a multicohort study

Serial oxygenation measurements may describe early risk in acute hypoxemic respiratory failure (AHRF), but incremental prediction beyond an initial PaO₂/FiO₂ (P/F) ratio is uncertain. We studied operationally defined AHRF in MIMIC-IV and replicated the trajectory framework in eICU-CRD. Mean P/F during 0–6 h (≥ 150 versus < 150 mmHg) and change during the second window ( ≥ + 30 versus < + 30 mmHg) defined four groups. Mortality prediction used a 12-hour landmark and a common MIMIC-IV sample, with repeated five-fold cross-validation, C-statistics, Brier scores, integrated Brier scores, and calibration. Among 14,322 eligible MIMIC-IV patients, 3,358 had paired early P/F measurements; 651 died within 28 days. Mortality was 11.6%, 14.9%, 23.6%, and 43.6% in groups A–D. Of 3,345 patients surviving to 12 h, 3,051 had complete comparison variables. The out-of-fold C-index was 0.624 for initial P/F, 0.647 for the 0–6-hour mean, and 0.666 for mean P/F plus change. Adding change to a model containing age, sex, vasopressor records, PEEP, and mean P/F changed the C-index from 0.703 to 0.706, with a small reduction in Brier score. In eICU-CRD, 12,141 patients had paired measurements and 12,020 had known hospital outcomes; the group mortality gradient was reproduced. Early P/F trajectory described mortality strata across two ICU databases. Continuous change provided modest additional predictive information, while incremental discrimination beyond clinical variables and early mean P/F remained uncertain. Not applicable.