Improving retrospective ARDS case-finding using a simple 72-h physiologic persistence rule

Abstract Background Retrospective studies frequently use single-time-point Berlin physiologic criteria (PaO2/FiO2 < 300 mm Hg plus positive end-expiratory pressure ≥ 5 cm H2O) to identify acute respiratory distress syndrome (ARDS). However, transient hypoxaemia is common among critically ill patients and often does not represent ARDS. Using these screens may overestimate ARDS prevalence and yield a cohort inconsistent with those seen in clinical trials. We sought to determine whether a 72-h persistence criterion for hypoxaemia improves the accuracy of ARDS case identification and evaluate the additional case-finding value of radiology keyword searches and ICD-codes. Methods We conducted a retrospective cohort study using the MIMIC-IV database (2008–2019) for derivation and a UK ICU dataset (Imperial College Healthcare National Health Service Trust, 2009–2024) for external validation. All patients meeting Berlin physiologic criteria for at least 72 h were identified. From MIMIC-IV, we randomly selected 2000 patients who met 72-h persistence criteria for expert adjudication based on detailed review of clinical notes, imaging, and echocardiography, classifying them as ARDS, non-ARDS acute hypoxaemic respiratory failure, or possible ARDS. Sensitivity analyses with shorter durations (≥ 24 and ≥ 48 h) were performed. Diagnostic performance of radiology keyword searches and ARDS-specific ICD-9/10 codes were compared to expert adjudication. Results Of 18,621 patients who ever met physiologic criteria, 3940 met the 72-h persistence threshold. In a random sample of 2000 from this 72-h MIMIC-IV cohort, expert adjudication identified ARDS in 49.7% (95% CI, 48–52%); in the external UK validation cohort, 56% (95% CI, 46–66%) were adjudicated as ARDS. ARDS prevalence significantly declined with shorter persistence requirements: 21% after 48 h, 8% after 24 h, and 6% with single isolated measurements. Within the 72-h persistence criterion enriched sample, the highest performing radiology keyword search set provided limited sensitivity (49%) and moderate specificity (76%), whereas ICD codes had higher sensitivity (76%) but low specificity (47%). Conclusions Berlin physiologic criteria alone were inadequate for retrospective ARDS identification. A ≥ 72-h persistence rule improved cohort enrichment but did not define ARDS, with substantial residual misclassification remaining after physiologic screening. Persistence should therefore be viewed as a pragmatic enrichment strategy rather than a definitive retrospective ARDS label.