TimeDiff

Reliable Generation of Privacy-preserving Synthetic EHR Time Series via Diffusion Models

This repository contains the code implementation for TimeDiff, a diffusion model designed for the generation of mixed-type time series data in EHR. Our work is accepted by the Journal of the American Medical Informatics Association (JAMIA).

How to Run TimeDiff

• Use requirements.txt to install dependencies.
• etdiff_train.py is the script to initialize model training. After training is complete, it will automatically generate synthetic EHR time series data and store it in the designated directory set by the user. This directory can be specified with --check_point_path command line argument.
• The specific implementation of the model can be found at models/ETDiff.

Data Pre-processing

Please see preprocess directory for Jupyter notebooks and Python scripts on how we preprocess EHR data for model training and evaluation.

Baselines

NOTE: Please set PYTHONPATH correctly to the root of this repository in order to run the baseline models.

Dependencies for the Baselines

We recommend using independent environments for each of the baseline models, as some of them have distinctively different requirements (one example is the version of Python for tensorflow 1.X). Please refer to the code repository for the baselines and Appendix A.4.4 in our paper for more details.

• EHR-M-GAN: run baselines/ehrmgan_train.py; implementation is stored at models/ehr_m_gan.
• DSPD/CSPD: run baselines/dspd_train.py; implementation is stored at models/tsdiff; to use CSPD rather than DSPD, use continuous diffusion for --diffusion argument.
• GT-GAN: run baselines/gtgan_train.py; implementation is stored at models/gt_gan.
• TimeGAN: run baselines/timegan_train.py; implementation is stored at models/time_gan.
• RCGAN: run baselines/rcgan_train.py; implementation is stored at models/rc_gan.
• C-RNN-GAN: run baselines/crnngan_train.py; implementation is stored at models/crnn_gan.
• T-Forcing/P-Forcing: run baselines/pt_forcing_train.py; implementation is stored at models/p_or_t_forcing; passing the command line argument -adversarial uses P-Forcing. Otherwise, T-Forcing is used.
• HALO: please refer to baselines/halo_preprocess.py and the descriptions provided by the HALO authors in their manuscript to preprocess the input data. Run baselines/halo_train.py to train the model and baselines/halo_generate.py for sampling. Implementation is stored at models/halo.

Evaluation Metrics

• eval_samples.py is the script responsible for running all the evaluation metrics discussed in the paper; to run the code, pass the intended metric to use via --metric command line argument, as well as the training data (--train_path), testing data (--test_path), and synthetic data (--sync_path).
• Implementations for data utility metrics can be found at evaluate/utility.py.
• Implementations for privacy risk metrics can be found at evaluate/privacy.py.

Citation

@article{tian2024reliable,
  title={Reliable generation of privacy-preserving synthetic electronic health record time series via diffusion models},
  author={Tian, Muhang and Chen, Bernie and Guo, Allan and Jiang, Shiyi and Zhang, Anru R},
  journal={Journal of the American Medical Informatics Association},
  pages={ocae229},
  year={2024},
  publisher={Oxford University Press}
}