Fine-Grained Forecasting of COVID-19 Trends at the County Level in the United States

Architecture

A novel time-series deep learning-based approach to predict short-term infection trends related to COVID-19 in the US Counties. This state-of-the-art forecasting models consists of bidirectional LSTM deep learning structure and county clusters to adapt to sudden dynamic changes, efficiently improving the learning effectiveness of relevant features. This research offers a valuable short-term epidemic prediction framework to aid governments in formulating public health policies and curbing disease transmission.

Installation

To install necessary library packages, run the following command in your terminal:

pip install -r requirements.txt

Usage

• Clone the repo to your project folder by using the following commend:

  git clone https://github.com/kleelab-bch/FIGI-Net

• Prepare the dataset as Excel file and copy to the Data folder.

• Follow the order of codes (in the src folder)

  • Run 1_Temporal_Clustering.py to obtain the cluster labels of US counties.
    • The Clustering labels will be saved to a new custom sheet in the original Excel file.
  • Then run 2_FIGINet_Prediction.py for model training and result forecasting.
    • If the user would like to use uni-pretrained models to train FIGI-Net models , please set the parameters Use_Uni_Pretrained to True and Use_Grained_Pretrained to **False**.
    • If the user would like to train all models from scratch, please set the above parameters to False.

• The forecasting results will be generated in Results folder

Note

• All the Covid-19 Confirmed Data of US Counties are from Center for Systems Science and Engineering (CSSE) at Johns Hopkins University.
• The lib folder includes all dependencies required for the FIGInet workflow.
• All trained models are saved to the Model folder.
• The raw error data, consisting of the RMSE and RRMSE values of evaluated models, have been added to the Data folder for comparison and analysis.

Citation

@article{Song2025FIGINet,
  author       = {Song, Tzu‑Hsi and Clemente, Leonardo and Pan, Xiang and Jang, Junbong and Santillana, Mauricio and Lee, Kwonmoo},
  title        = {Fine‑grained forecasting of COVID‑19 trends at the county level in the United States},
  journal      = {npj Digital Medicine},
  volume       = {8},
  number       = {1},
  pages        = {204},
  year         = {2025},
  doi          = {10.1038/s41746-025-01606-1},
  url          = {https://www.nature.com/articles/s41746-025-01606-1}
}

Contact

If you have any question about the code or paper, please contact Tzu-Hsi.Song@childrens.harvard.edu