Blood-gas history and arterial oxygen target classification after acute brain injury: a temporal and multicentre external validation study

Abstract Background Oxygen management after acute brain injury (ABI) is guided by arterial oxygen tension (PaO₂), whereas bedside monitoring relies on pulse oximetry (SpO₂). Every intensive care unit (ICU) stores a blood-gas history that is rarely reused. We tested whether this history improves PaO₂ target classification beyond SpO₂ across eras and hospitals and supports reliable in-target calls. Methods In this retrospective study of adults with ABI (MIMIC-IV 3.1, eICU 2.0), arterial blood gases from the first 72 ICU hours were paired with SpO₂ recorded ≤ 10 minutes earlier; only results available before each draw counted as history. PaO₂ categories were < 80, 80–120 and > 120 mmHg, the middle range reflecting consensus for mechanically ventilated ABI patients. Penalised multinomial models using current information (SpO₂, age, ICU time), blood-gas history or both were developed in MIMIC-IV and evaluated without refitting in patient-disjoint later-era MIMIC-IV and eICU data. The primary outcome was the external Brier score difference; a prespecified ≤ 5% error gate governed automated in-target calls. Results Cohorts comprised 2,070 MIMIC-IV patients (4,283 pairs) and 2,373 eICU patients (7,731 pairs). In 4,877 external measurements with history (1,493 patients, 103 hospitals), adding blood-gas history improved the Brier score from 0.5753 to 0.5289 (− 0.0464; 95% CI − 0.0536 to − 0.0390), an 8.1% relative reduction. Improvement was replicated in later-era MIMIC-IV, was directionally consistent in 21 of 22 larger hospitals and persisted in eICU for blood gases > 24 hours old. Post hoc adjustment for recorded inspired oxygen fraction, positive end-expiratory pressure and ventilation status yielded a similar difference (− 0.0486; −0.0556 to − 0.0413). In-target classification, however, failed: 46.6% of the combined model's external calls were wrong, two in five eICU measurements with SpO₂ 94–98% had PaO₂ <80 mmHg, and no model passed the error gate. Conclusions Blood-gas history stored in ICU databases provides reproducible incremental information for PaO₂ target classification after ABI, including after adjustment for recorded respiratory support, but the evaluated models could not reliably confirm in-target PaO₂. These findings support displaying the latest blood gas and its age alongside SpO₂, retaining arterial confirmation, and requiring external validation and prespecified error gates before deploying database-derived tools.