Reproducibility Code and Model Configuration for a Nurse-Centered Early-Warning Framework for Diabetic Ketoacidosis

Abstract

This repository provides the reproducibility materials associated with the research manuscript entitled: “Simulation-Based Development, Stress Testing, and External Physiological Benchmarking of a Nurse-Centered Early-Warning Framework for Diabetic Ketoacidosis.” The repository is intended to support transparency, reproducibility, and independent verification of the model-development and evaluation procedures described in the manuscript. The reproducibility materials include the primary model-development code, peer-review analysis code, and supporting configuration and documentation files. Model Development and Reproducibility The study evaluated machine-learning approaches, including XGBoost and logistic regression, for the development and assessment of a nurse-centered early-warning framework for diabetic ketoacidosis. The following methodological specifications are documented in the accompanying materials: 1. Hyperparameter Selection Final model hyperparameters were fixed in the configuration file config_final.yaml before the final version 4 (v4) analysis. No automated hyperparameter tuning was performed on the independent test set. 2. Class Imbalance Handling For XGBoost, the scale_pos_weight parameter was set to 43.902, based on 2,208,444 negative and 50,304 positive training time points. Logistic regression was implemented using balanced class weights. 3. Missing Data Handling Primary feature engineering required a complete 60-minute observation history. For missingness stress testing, missing observations were handled using within-episode forward filling for a maximum of 15 minutes, followed by median replacement before recalculating the model features. 4. Software Environment and Reproducibility The accompanying reproducibility materials document the software environment, package versions, random seeds, and model configurations used in the final analyses. 5. Code Availability The repository provides the primary model-development and peer-review analysis code corresponding to Supplementary Files S2 and S3. These materials are intended to facilitate independent examination of the analytical procedures and reproducibility of the reported findings. Data Availability The availability of underlying datasets is subject to the applicable ethical approvals, data governance requirements, and restrictions described in the associated manuscript. Associated Publication Simulation-Based Development, Stress Testing, and External Physiological Benchmarking of a Nurse-Centered Early-Warning Framework for Diabetic Ketoacidosis.