Forecasting with irregularly measured blood glucose with deep learning

Repository for the paper:

Continuous time recurrent neural networks : overview and application to forecasting blood glucose in the intensive care unit

Data setup

MIMIC-IV

As we are not data custodians we cannot publicaly share the MIMIC-IV data used. However, it is available to those with credentialed access to physionet.org. Credentialed access can be requested through your physionet account.

We used dbt to connect to the Google Bigquery MIMIC-IV database that can be autopopulated through the physionet interface.

https://docs.getdbt.com/reference/warehouse-setups/bigquery-setup

After setting up dbt run:

cd scripts/data/mimic4glucose
dbt run

Then run the notebook: scripts/data/setup_mimic_data.ipynb

Simulations

After installing torchctrnn run the notebook: scripts/data/setup_simulations.ipynb

Notable dependencies

• python 3.9+
• dbt
• db-dtype
• pytorch
• torchctrnn: pip install https://github.com/oizin/torchctrnn/tarball/main
• numba
• properscoring

Repeating the experiments

The full analysis is reproducible as follows:

1. Run the bash scripts:

./run_simulation_experiments.sh
./run_mimic_experiments.sh

2. Run the evaluation notebooks

• scripts/results/simulation_data_size.ipynb
• scripts/results/simulation_all_5000.ipynb
• scripts/results/mimic_results_from_predictions.ipynb

The following is an example of running a single experiment

python train.py --model=LinearModel --logfolder=results --test --nfolds=1 --seed 1 --data mimic