Gluformer: Transformer-Based Personalized Glucose Forecasting with Uncertainty Quantification

This is the official implementation of the paper "Gluformer: Transformer-Based Personalized Glucose Forecasting with Uncertainty Quantification" (link).

---

Table of Contents:

• 1. Overview
• 2. How to work with the repository?
  • 2.1 Setting up the environment
  • 2.2 Running the experiment
  • 2.3 Downloading the data
  • 2.4 Running the scripts
• 3. How to cite

---

1. Overview

The code is organized as follows:

• cache:
  • visualize_* contains code for reproducing plots from the paper.
• gludata
  • data folder containing the data pre-processing tools.
  • data_loader.py provides the implementation of PyTorch Dataset for the data.
• gluformer provides our model implementation.
• trials contains outputs from the experiments on the real / synthetic data sets.
  • trials.txt provides commands to run the experiments.
• utils contains common tools for model training / evaluation.
• experiment.ipynb provides an example on the synthetic data of how to train and evaluate our model.
• model_* scripts for model training and evaluation respectively that can be run from the command line.

Additionally, we provide environment.yaml that gives a snpshot of our conda build for reproducibility.

2. How to work with the repository?

The repository provides a self-contained code to repoduce the results from the paper on the glucose and synthetic data sets. Below, we outline some futher instructions.

2.1 Setting up the enviroment

We recommend use conda to create a virtual environment for this project. Once you have pulled the code and installed conda on your system, you need to run the following command from the root (repository) folder: conda env create -f environment.yml. Once the necessary packages are installed, you can activate your environment by running conda activate gluformer.

2.2 Running the experiment

We suggest to start the explortion of our model by running the model on the synthetic data. The code for this is provided in the experiment.ipynb notebook. The code is largely self-contained as it gives the implementation of both the training and the evaluation loops. Additionally, the notebook contains the data generating function for the synthetic data. All of this is done to expose the use to the inner workings of the model and ease potential extensions to the other data sets.

2.3 Downloading the data

For the results obtained in the paper, we use a publicly available CGM data set provided here. We recommend downloading the data set directly from the authors' repository and using our script gludata/data/split.py to split the data. You can visualize and process the data at the same using the gludata/data/view.ipynb notebook. If you have trouble with the downloading or pre-processing the data, do not hesitate to raise an issue.

(Update March 20, 2023) For simplicity, we provide the data in our repository. The full data set is UM_data.pkl. The data corresponds to merging the processed_cgm_data_train.pkl, processed_cgm_data_validation.pkl, and processed_cgm_data_test.pkl in the source repository. The data comes under the CC BY-NC-SA license, which implies that:

• You will no attempt re-identification.
• You will contact the University of Michigan if identifiers are detected.
• You will not redistribute or resell the data.
• Data ownership remains with the University of Michigan.
• Requirements survive changes in ownership of entity.

2.4 Running the scripts

We provide model_train.py and model_eval.py scripts that can be run from the command line and give an implementation of the training and evaluation loops respectively. Both scripts expect to be run from the root (repository) folder and have all dependencies (specified in the environment.yaml file) installed. For an example of what parameters each script takes, see the trials.txt file in the trials/ folder.

3. How to cite

@inproceedings{sergazinov2022gluformer,
    title={Gluformer: Transformer-Based Personalized Glucose Forecasting with Uncertainty Quantification},
    author={Renat Sergazinov and Mohammadreza Armandpour and Irina Gaynanova},
    booktitle={{arXiv}},
    year={2022},
}