momoyamasawa/accuratebg

Patient-specific blood glucose prediction using deep learning, considering the challenges of "small dataset" and "data imbalance"

Accurate BG

Repository for paper Deep transfer learning and data augmentation improve glucose levels prediction in type 2 diabetes patients.

Installation

We recommend you create a virtual environment via Anaconda or Pyenv, then activate the virtual environment and run

>> make dev

We used tensorflow 1.15.0.

Train and test for the OhioT1DM dataset

To test for the OhioT1DM dataset, create a folder named data at the root directory

>> mkdir data/

Then, download the public dataset OhioT1DM here, and move the unzipped folder OhioT1DM into data. (Note that to use the dataset, you need to apply for it with the link provided and we are unfortunately not able to provide the dataset for you.)

To train and test for the OhioT1DM dataset, with our optimal configuration, run

>> cd accurate_bg
>> python3 ohio_main.py --epoch 150

the default prediction horizon is 6, equivalently 30 min. To adjust prediction horizon to 1hr, modify the last line of command in the code block above with

>> python3 ohio_main.py --epoch 150 --prediction_horizon 12

Reference

If you find this repo or our work helpful, we encourage you to cite the paper below.

@article{deng2021deep,
  title={Deep transfer learning and data augmentation improve glucose levels prediction in type 2 diabetes patients},
  author={Deng, Yixiang and Lu, Lu and Aponte, Laura and Angelidi, Angeliki M and Novak, Vera and Karniadakis, George Em and Mantzoros, Christos S},
  journal={NPJ Digital Medicine},
  volume={4},
  number={1},
  pages={1--13},
  year={2021},
  publisher={Nature Publishing Group}
}