README.md

Predictive Performance Precision Analysis in Medicine: Identification of low-confidence predictions at patient and profile levels (MED3pa I)

This repository is to generate results of the article "Predictive Performance Precision Analysis in Medicine: Identification of low-confidence predictions at patient and profile levels (MED3pa I)", published in the Journal of the American Medical Informatics Association (JAMIA): https://doi.org/10.1093/jamia/ocag034.

The MED3PA package used in this work is available at this link: https://github.com/MEDomicsLab/MED3pa

The MED3PA package can be installed from pypi: https://med3pa.readthedocs.io/en/latest/installation.html

1. Data availability

This study uses three types of datasets:

• Simulated Data A synthetic dataset is generated in this repository, using the 'generated_simulated_dataset.py' script in the datasets/simulated_dataset folder.
• Public Clinical Datasets (In-hospital mortality task)
  • eICU Collaborative Research Database (eICU) (v2.0) – Requires credentialed access through PhysioNet.
  • MIMIC-IV (V1.0) – Also requires credentialed access through PhysioNet.
    Users must complete the required training and data use agreements to access these datasets.
• Private Clinical Dataset (One Year mortality task, POYM) Due to regulations safeguarding patient privacy, this dataset cannot be shared. However, a synthetic dataset is publicly available on https://zenodo.org/doi/10.5281/zenodo.12954672.

2. Study recreation

To reproduce experiments and figures from the paper:

1. Install requirements First install the requirements under Python 3.12.4 as following:

$ pip install -r requirements.txt

2. Generate simulated Data

$ python -m datasets.simulated_dataset.generate_simulated_dataset

3. Run experiments

$ python -m experiments.simulated_dataset

This will generate results in the results folder.