Replicating "Design and implementation of a deep recurrent model for prediction of readmission in urgent care using electronic health records" (Zebin and Chaussalet, 2019)

Note: To run this project, you will need a copy of the MIMIC III dataset located on the machine you wish to run. The MIMIC III dataset is publicly available through credentialed access on Physionet.org.

This project requires:

Additionally, you will need to install Jupyter to run the interactive notebooks.

If you already have a suitably sized subset of the MIMIC III dataset, you can skip this step.

By running the data_sampling/datasampler.py script, you can generate a subset of the full MIMIC III dataset that will more easily run on your machine. We used a sample size of 10000 ICU stays.

Note: you will need to edit this script with the correct path to the copy of the MIMIC III dataset that you downloaded on your machine.

Running the data sampler:

python3 data_sampling/datasampler.py

The data must be preprocessed and stored into binary files before it is used by the models. The following scripts from the data_cleaning directory should be run before launching the interactive notebooks:

python3 data_cleaning/preprocessing_mp.py
python3 data_cleaning/events_to_list.py

The training of the various models is performed in Jupyter interactive notebooks. To run one of these models, launch the Jupyter server with file of the model you wish to train:

# Open LSTM+CNN Model interactive notebook
jupyter notebook source/lstm-cnn-model.ipynb

# Open LSTM Model baseline interactive notebook
jupyter notebook source/lstm-model.ipynb

# Open Logistic Model baseline interactive notebook
jupyter notebook source/logistic-model.ipynb

The results history of the models are stored as pickle binary files in the same directory.

This project repository is organized into the following sub-folders:

• data: binary working folders of preprocessed data (MIMIC-III data must be added according to instructions above)
• demo_data: binary working folders of preprocessed data from publicly available demo data used for model development
• data_sampling: data subset sampling source code
• data_cleaning: data preprocessing source code
• project_summary: tables and figures summarizing reproducibility results
• source: interactive notebooks for each model and utility classes

T. Zebin and T. J. Chaussalet, "Design and implementation of a deep recurrent model for prediction of readmission in urgent care using electronic health records," 2019 IEEE Conference on Computational Intelligence in Bioinformatics and Computational Biology (CIBCB), Siena, Italy, 2019, pp. 1-5, doi: 10.1109/CIBCB.2019.8791466.