Pre-registered analysis plan: How much do real-time early-warning models add beyond the current state in the surgical ICU? A two-database evaluation (MIMIC-IV, eICU-CRD)
Pre-registered study protocol and statistical analysis plan for a retrospective evaluation of real-time early-warning models in adults admitted to surgical-type intensive care units. The study compares a model that uses recent physiological history with an identically tuned model that uses only a snapshot of the current state. Models are developed in MIMIC-IV v3.1 and evaluated externally in eICU-CRD v2.0. Primary outcome: recorded new-onset sustained hypotension within 8 hours. Secondary outcomes: hypotension or vasopressor action, respiratory deterioration, and acute kidney injury (KDIGO stage 2 or higher). Primary analysis: difference in AUROC between the two models, with a paired hospital-cluster bootstrap confidence interval. Key secondary analyses: within-hospital discrimination and early detection of hypotension episodes at an equal alert burden. The plan was registered before any data were extracted or analysed.
Authors
- Nhat An Nguyen (ORCID: https://orcid.org/0009-0008-9508-9613)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-12
- DOI
- https://doi.org/10.5281/zenodo.24148692
- Primary Topic
- Machine Learning in Healthcare
- Type
- article
- Field-Weighted Citation Impact
- 0.00