Citation

• Article: https://link.springer.com/article/10.1007%2Fs41666-019-00059-y

Please cite this work if you find this repository useful for your research:

Martinsson, J., Schliep, A., Eliasson, B. et al. J Healthc Inform Res (2019). https://doi.org/10.1007/s41666-019-00059-y

Prerequisites

The code is designed to be run on the OhioT1DM Dataset. So to use it the xml_path in e.g. the example experiment YAML configuration need to point to the path on disk where the XML data files are. E.g., change "/home/ubuntu/ohio_data/OhioT1DM-training/" to point to Ohiot1DM-training folder containing the XML files for the ohio dataset.

It would of cource be possible to write a new dataset module which loads the data into the required format and train the models on other data as well.

Installation

$> chmod +x setup.sh
$> ./setup.sh

Running an experiment

Note that this is designed to run on the Ohio Diabetes dataset. You need to explicitly state the absolute file path to the XML file of the patient you want to train the model for in the experiment configuration file (YAML file).

Except for that, everything should run out of the box.

$> chmod +x run.py
$> ./run.py --file experiments/example.yaml -m train

All results are collected and stored in the 'artifacts' directory. To visualize the training session you can run

$> tensorboard --logdir artifacts/<artifacts-path>

and fire up tensorboard.

Reproduce final results

$> sh run_final_experiments.sh
$> sh evaluate_final_experiments.sh
$> python create_table.py

Reproduce plots

# Run the hyperparameter search
$> python run.py --file=experiments/all_nb_lstm_state_nb_past_steps_search.yaml -m train
# Evaluate the trained models
$> python run.py --file=experiments/all_nb_lstm_state_nb_past_steps_search.yaml -m ealuate
# Hyperparam search plots
$> python plot_parameter_search.py artifacts/all_nb_lstm_states_nb_past_steps/

The plots will be in the working directory.

# Surveillance error grid plots and prediction plots
$> sh run_final_plots.sh

The plots will be in the artifacts folders.

Versions

To reproduce the results in Automatic blood glucose prediction with confidence using recurrent neural networks revert to commit: a5f0ebcf45f87b63d118dcad5e96eb505bb4269a and follow the README.