These are descriptive additions after the v1 analysis lock, prepared for manuscript reporting; no models were refitted, no thresholds were changed, and the primary endpoint, cohorts, pairing window and evaluation partitions were unchanged. All laboratory/monitoring data came from existing cached derived tables. The only additional raw-data read was the small MIMIC patient metadata table to recover recorded sex.

Patient-level summaries count each person once within their frozen partition. MIMIC uses the earliest primary-paired ICU admission in within-patient shifted chronology. eICU lacks complete cross-admission dates: a deterministic representative admission is selected by ascending hospital discharge year, hospital episode identifier, hospital-relative ICU admission minute and ICU stay identifier. This is not certified as the chronological first admission. Admission-level age, recorded sex and code flags are reported at this representative admission. Age is capped at90 years, matching the frozen models and the eICU >89 category. Overlapping ABI flags are retained and need not sum to100%. Sex describes the recorded database category, not a separately ascertained biological characteristic or gender identity.

Patient history groups are disjoint: ever versus never having an eligible available-history paired measurement anywhere among that patient's included pairs. All included pairs of such patients remain in the corresponding patient-group description. Measurement-level history strata instead classify each individual draw as eligible or ineligible; a patient can contribute to both measurement strata. The unavailable-history stratum is not synonymous with first measurement: it also includes chronological prior pairs that were not demonstrably available. All-pair class distributions include first/no-history draws. No baseline significance tests, causal comparisons or new outcome models were performed.

The flow starts from the existing adult stay cache, so it does not claim new enumeration of paediatric/unknown-age exclusions from the full database. Stay-based screening groups are mutually exclusive, but patient counts across these groups can overlap through multiple admissions. Arterial-reference and pairing losses are measurement-key counts; patient counts must not be subtracted as if all branches were disjoint. Accepted samples without a qualifying pre10 SpO2 pair are shown separately. MIMIC boundary-era patients are descriptive only, not silently reassigned to training or temporal validation. The authors confirmed review of the code-based eligibility rules and variable definitions during manuscript finalisation. The observed-code and label inventories remain aggregate descriptions, not patient-level clinical adjudication; unused review-workflow columns were removed from the submission copies without changing codes, flags or counts.
