Conditional Augmentation GAN (CA-GAN)

This is the repository for the Conditional Augmentation GAN (CA-GAN).

Details about the project can be found in our paper "Generative AI Mitigates Representation Bias Using Synthetic Health Data".

Link to the paper: Generative AI Mitigates Representation Bias Using Synthetic Health Data

Link to our previous work on CA-GAN: [2210.13958] Mitigating Health Data Poverty: Generative Approaches versus Resampling for Time-series Clinical Data

If you use this repository, please cite:

@article{micheletti2023generative,
 title={Generative AI Mitigates Representation Bias Using Synthetic Health Data},
 author={Micheletti, Nicolo and Marchesi, Raffaele and Kuo, Nicholas I-Hsien and Barbieri, Sebastiano and Jurman, Giuseppe and Osmani, Venet},
 journal={medRxiv},
 pages={2023--09},
 year={2023},
 publisher={Cold Spring Harbor Laboratory Press}
}

Data Access

The data for our project came from the MIMIC-III database. We cannot provide the data directly, but access to the database can be requested through PhysioNet (https://physionet.org/content/mimiciii/1.4/).

We chose two datasets extracted from the MIMIC-III database. The inclusion and exclusion criteria are described in the literature:

• Acute hypotension: Interpretable Off-Policy Evaluation in Reinforcement Learning by Highlighting Influential Transitions GitHub - dtak/interpretable_ope_public

• Sepsis: The Artificial Intelligence Clinician learns optimal treatment strategies for sepsis in intensive care | Nature Medicine AIClinician / AIClinician · GitLab

To demo this code without access to Physionet, please refer to Appendix B of our paper to recreate a sample dataset.

Usage

Get a copy of this project and set it up on your local machine.

Install the required packages (e.g. pip install -r requirements.txt).

The folders "Hypotension" and "Sepsis" are independent of each other. They represent the two case studies on which we applied CA-GAN. Both folders contain the code needed to:

• retrieve and preprocess data from the MIMIC-III database to obtain the real data used to train our model.
• train CA-GAN on the real data
• use the trained model to generate the synthetic data
• evaluate the quality of the data generated

Support

For support or inquiries, please contact [email protected] and [email protected]

Acknowledgment

This project is based on Kuo, Nicholas I-Hsien, et al. "The Health Gym: synthetic health-related datasets for the development of reinforcement learning algorithms." Scientific Data 9.1 (2022): 693. (Copyright (c) 2022. by Nicholas Kuo & Sebastiano Babieri, UNSW.)

License

This project is licensed under the MIT License - see the LICENSE file for details.