Patient cohorts for deep learning-based early warning in the intensive care unit

Paper

This repo uses the ricu R package to derive patient cohorts for prediction tasks from the following four intensive care databases:

• AUMCdb
• HiRID
• eICU
• MIMIC IV

If you use this code in your research, please cite the following publication:

@article{rockenschaub2023generalisability,
      title={Generalisability of deep learning-based early warning in the intensive care unit: a retrospective empirical evaluation}, 
      author={Patrick Rockenschaub and Adam Hilbert and Tabea Kossen and Falk von Dincklage and Vince Istvan Madai and Dietmar Frey},
      year={2023},
      eprint={2303.15354},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

This paper can be found on arxiv: https://arxiv.org/abs/2303.15354

Acknowledgements

The code in this repository heavily utilises the ricu R package, without which deriving these cohorts would have been much more difficult. If you use the code in this repo, please go give their repo a star :)

To replicate the cohorts:

Step 0: Environment and Prerequisites

Run the following commands to clone this repo:

git clone https://github.com/prockenschaub/icuDG-preprocessing.git
cd icuDG-preprocessing/

All data extractions were run using R 4.2.2 on an Apple M1 Max with Ventura 13.2.1. An renv lock file was created to install all necessary package dependencies. To recreate the environment, start an R session and call the following commands:

install.packages("renv")
renv::restore()

Step 1: Load the datasets into ricu

In order to access the full datasets through R, you need to download them and make them available to ricu. Please follow the instructions given by the ricu package: ?ricu::import_src.

For quick experimentation, ricu comes with two demo datasets: mimic.demo and eicu.demo. These are small, openly available subsets of mimic and eicu that allow for easy prototyping. They should have been installed by renv. If they aren't, please see the respective Github pages here and here.

Step 2: Generate the Cohorts

This repository currently allows for the extraction of three distinct prediction tasks:

1. ICU mortality after 24 hours: mortality.R
2. Acute Kidney Injury within the next 6 hours: aki.R
3. Sepsis within the next 6 hours: sepsis.R

All three tasks rely on a shared data cleaning provided in base_cohort.R, which defines and stores a subset of patients in each dataset with sufficient data quality. base_cohort.R therefore needs to be called before any of the task-specific cohorts can be generated.

Once base_cohort.R was run, data for a single task from a single dataset can be extracted via:

Rscript "mortality.R" --src mimic_demo

where mortality.R should be replaced with the task file of interest and mimic_demo with the database of interest (one of mimic_demo, eicu_demo, aumc, hirid, eicu, mimic, miiv). Data can be extracted from all datasets simultaneously via bash gen_cohort.sh mortality.

The output directory for the extracted data can be set in config.json

License

This source code is released under the MIT license, included here.