Type 2 Diabetes · MI · AF × Age > 70 — ICU Mortality Miscalibration

Cohort inclusion criteria

• Adult ICU admissions in MIMIC-IV v2.2 (PhysioNet credentialed access).
• First ICU stay per hospital admission only.
• Length of stay ≥ 6 hours (excludes transient stays not reflective of ICU care).
• Age computed at the start of the ICU stay; stratified at age > 70 for the three primary findings, and age < 50 for the COPD comparator.
• No exclusion based on race, ethnicity, sex, insurance, or admission source.

ICD code stacks (per condition)

• Type 2 diabetes: ICD-9 250.x0, 250.x2; ICD-10 E11.*. Excludes type 1 (250.x1, 250.x3, E10.*) and gestational (O24.4).
• Acute myocardial infarction: ICD-9 410.*; ICD-10 I21.*, I22.*. Index admission must be the primary diagnosis or first secondary diagnosis with billed troponin elevation.
• Atrial fibrillation: ICD-9 427.31; ICD-10 I48.0, I48.1, I48.2, I48.91. Includes paroxysmal, persistent, chronic, and unspecified.
• Pulmonary embolism: ICD-9 415.1*; ICD-10 I26.*.
• Seizure: ICD-9 345.*, 780.39; ICD-10 G40.*, R56.9.
• COPD: ICD-9 491.*, 492.*, 496; ICD-10 J44.*.

Outcome definition

In-hospital mortality. Patient discharge disposition coded as expired or death, evaluated at the hospital-admission level. Patients transferred to hospice from inpatient care are not counted as in-hospital deaths in the primary analysis (sensitivity analysis below treats hospice transfer as an event).

Analytical plan

1. Compute observed in-hospital mortality per stratum.
2. Compute Wilson score 95% confidence intervals on each proportion (continuity-corrected).
3. Compute relative divergence from the published consensus estimate as (observed − published) / published × 100.
4. Bootstrap risk difference (10,000 resamples) for the divergence point estimate on each stratum.
5. For confirmation: re-run identical pipeline on eICU-CRD v2.0 (PhysioNet credentialed access). Confirmation requires the divergence direction to replicate and the eICU 95% CI to exclude the published consensus value.

Pre-specified sensitivity analyses

• Hospice-as-event: count hospice transfer as in-hospital death.
• 30-day mortality variant: use 30-day post-admission mortality in place of in-hospital, where MIMIC-IV linkage permits.
• First-ever vs. recurrent: stratify by whether the index condition is first-ever or recurrent (where prior admission data exists).
• Severity adjustment: repeat with SOFA at 24 h and APACHE-IVa included as covariates; report effect on point estimate.
• Era effect: split MIMIC-IV by admission year (2010–2015 vs. 2016–2023) to test for temporal drift in observed mortality.

Decision rule (pre-specified)

A finding is labeled Confirmed only if all of: (a) MIMIC-IV divergence ≥ 50% in absolute relative terms, (b) eICU-CRD replication of direction, (c) eICU-CRD 95% CI excludes the published consensus, and (d) novelty score ≥ 0.75 from the engine's adversarial review (Advocate / Adversary / Arbiter unanimous). Findings failing any criterion remain in the Exploratory tier, single-dataset and OSF-timestamped, pending external replication.