Figure 1. Cohort construction and evaluation partitions.
The flow shows adult ICU screening, acute brain injury (ABI) code candidates after documented special-background exclusions, accepted arterial-reference oxygen measurements, primary pulse oximetry–blood-gas pairs and available-history paired measurements. Current SpO₂ was the most recent valid measurement at or before the blood-gas time proxy within 10 min, restricted to the first 72 h of each ICU stay. MIMIC used explicit arterial specimen documentation; eICU used arterial-named PaO₂ within the arterial blood-gas laboratory category with pressure units. Patient-level partitions kept all encounters together. MIMIC patients whose calendar-era intervals crossed 2020 contributed only to descriptive summaries. Partition entries show all paired patients and history-eligible measurements, whose patient denominators are 584, 135, 180 and 47 in training, calibration, later-era and boundary groups, respectively. Exclusion-branch patient counts can overlap and are not represented as sequential patient losses. Two eICU centre counts distinguish all primary pairs from the matched historical comparison.

Figure 2. Probability-classification increment and the error–coverage relationship.
(a) Pair-weighted three-class Brier scores for the prior-only model (P), current-information model (B) and current-plus-prior-information model (C), evaluated on identical history-eligible measurements: 423 measurements from 180 patients in the MIMIC later-era set and 4,877 measurements from 1,493 patients in eICU. The sum-of-three-class Brier score was not divided by 3; lower values indicate better probability predictions. Intervals are 95% percentile intervals from 1,000 patient-cluster bootstrap replicates. (b) Descriptive eICU false in-target fractions versus coverage over all 101 probability thresholds from 0 to 1. The false in-target fraction is the proportion below 80 or above 120 mmHg among selected measurements; coverage uses all 4,877 history-eligible measurements. Nonempty selections with fewer than 30 measurements or 20 patients are hollow and disconnected; their denominators are printed. Zero-selection thresholds remain in the source table but cannot be plotted as an error fraction. The dotted line marks the prespecified 5% error criterion used as the deployment gate. This external curve was not used to select or revise thresholds. No model passed the gate in MIMIC calibration, so the locked selective policy released no in-target calls.

Figure 3. External calibration of the three oxygen-reference categories.
Calibration plots show observed versus mean predicted probabilities in fixed 0.1-wide probability bins for models P (prior information), B (current information) and C (current plus prior information) in 4,877 eICU history-eligible measurements from 1,493 patients. Categories are below 80, 80–120 inclusive and above 120 mmHg. The dotted diagonal denotes ideal calibration. Point size increases with the number of measurements in the bin; all occupied bins are retained. Lines connect ordered probability bins. Bin membership and patient counts are available in calibration_bins.csv. No external recalibration was applied. 

Figure S1. Historical-information age and incremental probability-classification performance.
The curves describe the per-measurement Brier loss difference C−B versus hours since the selected prior blood gas on a logarithmic time axis. Negative values favour model C. The descriptive restricted cubic spline used prespecified knots corresponding to 1, 6, 12 and 24 h on log(1+hours). Shading denotes 95% pointwise percentile bands from 1,000 patient-cluster bootstrap replicates. Cohorts contain 423 measurements from 180 patients (MIMIC later era) and 4,877 measurements from 1,493 patients (eICU). This characterization did not modify model predictions or determine a new time cutoff. The blood-gas sampling process was clinically selected; the curves do not establish a safe sampling interval or universal expiry time. 

Figure S2. Hospital-level variation in the historical-information increment.
Points show pair-weighted Brier differences C−B and 95% patient-cluster percentile intervals for eICU hospitals with at least 20 history-eligible patients; n labels denote patients, not measurements. Global patient resampling kept all of a patient's records together, including records across hospitals. Negative values favour C. Smaller hospitals remained in the overall analysis and were pooled for descriptive summaries rather than assigned unstable separate performance claims. The external matched comparison overall contains 4,877 measurements from 1,493 patients across 103 hospitals. Intervals are conditional on frozen models and the observed hospitals; they do not represent uncertainty from random sampling of a new hospital population. No multiplicity-adjusted hospital-specific efficacy claim is made.
