SyntheticDataFairness
The repository includes the code for metrics designed for evaluating fairness of synthetic healthcare datasets and how they are applied for fairness quantification. This repository is the official repository for the paper: The Problem of Fairness in Synthetic Healthcare Data.
Repository structure
- data: The folder includes data for two datasets. Atus is the American Time Use Survey dataset, both the derived real and synthetic data files. Mimic is the MIMIC-III dataset based on a past study for identifying the impact of race on mortality and includes only the synthetic dataset. Note that the synthetic datasets are generated using a Generative Adversarial Network (GAN) model called HealthGAN and are intended to not release any private information of the real datasets.
- scripts: The scripts include code snippets which are used in multiple other files or notebooks and hence, have been designed to be imported as functions.
- notebooks: The notebooks include code for plotting figures and calculating metrics on the datatsets.
- results: The results for the log disparity metric on synthetic datasets is compiled into CSV files included in this folder.
Data files description
- ATUS: ATUS dataset has both the real and synthetic files available in this repository. The real data file is called atus_train.csv and the synthetic file is called atus_train_synthetic_1.csv. The real and synthetic data are used based on the previously published paper Medical Time-Series Data Generation Using Generative Adversarial Networks.
- MIMIC: MIMIC dataset has only the synthetic file available. The synthetic file is called mimic_3_synthetic.csv. The synthetic data is used based on the previously published paper Generation and evaluation of privacy preserving synthetic health data.
Contacts
For questions, please reach out to Karan Bhanot (bhanok@rpi.edu or bhanotkaran22@gmail.com).