PostopAKI Transport: three-database analysis code and aggregate results for creatinine-defined learning targets and local adaptation

Abstract

This release contains analysis code and aggregate results supporting the manuscript “Creatinine-defined learning targets and local adaptation of perioperative acute kidney injury prediction models”. The study uses MIMIC-IV v3.1, INSPIRE v1.4.2 and MOVER EPIC records, with evaluable cohorts of 27,735, 44,194 and 8,890 hospitalizations, respectively. It examines how creatinine-based outcome construction affects model learning and external evaluation, and how imported source scores can support local learning with limited development data. Analyses include transport and recalibration, controlled input and training-label processing, event-count and structured controls, local adaptation at two patient budgets, and fixed-capacity event coverage and priority-list turnover. Version 2.0.0 includes 71 aggregate CSV files, an aggregate processing codebook, five main figures and one supplementary figure with reproduction code, data dictionaries, tested environment information, a result-to-code map, and configurable restricted-data analysis entry points. The public aggregate workflow rebuilds the figures and main transport table without clinical records. The documentation distinguishes fresh fitting, fixed-prediction evaluation and bounded checks against historical caches. Raw clinical records, patient identifiers, patient-level features or predictions, fitted model objects, private replay manifests and credentials are not included. Clinical-data analyses require independently authorized access through the database providers. Exact historical MIMIC-IV/INSPIRE partition and tied-value replay additionally depends on restricted manifests described in the archive; the public ZIP alone does not provide those inputs. A documented deterministic alternative supports a new analysis. This is a software and aggregate-results release, not a manuscript preprint or a deployable clinical prediction tool. Version 2.0.0 extends the earlier two-database release 1.0.0 (https://doi.org/10.5281/zenodo.23015065); that earlier release does not cover the present three-database analyses. Original software is licensed under the MIT License. Original documentation, figures and aggregate statistical reports are licensed under CC BY 4.0. These licenses apply to their respective components and do not alter the terms governing MIMIC-IV, INSPIRE or MOVER source data. This work was supported by the Traditional Chinese Medicine Research Project of Zhejiang Province (grant no. 2024ZL915), the Medical and Health Research Project of Zhejiang Province (grant no. 2025HY0970), and the Project of Ningbo Leading Medical & Health Discipline (discipline no. 2026-A14). The funders had no role in the study design, data analysis or interpretation, manuscript preparation, or decision to submit the manuscript.