Benchmarking the performance of group-fair and minimax-fair methods on two chest x-ray datasets, with an auxiliary investigation into label bias in MIMIC-CXR. For more details please see our CHIL 2022 paper.
Run the following commands to clone this repo and create the Conda environment:
git clone git@github.com:MLforHealth/CXR_Fairness.git
cd CXR_Fairness/
conda env create -f environment.yml
conda activate cxr_fairness
See DataSources.md for detailed instructions.
To reproduce the experiments in the paper which involve training grids of models using different debiasing methods, use cxr_fairness/sweep.py
as follows:
python -m cxr_fairness.sweep launch \
--experiment {experiment_name} \
--output_dir {output_root} \
--command_launcher {launcher}
where:
experiment_name
corresponds to experiments defined as classes in cxr_fairness/experiments.py
output_root
is a directory where experimental results will be stored.launcher
is a string corresponding to a launcher defined in cxr_fairness/launchers.py
(i.e. slurm
or local
).Sample bash scripts showing the command can also be found in bash_scripts/
.
The ERM
experiment should be ran first. The remaining experiments can be ran in any order, except JTT
should be ran after ERM
and after updating the path in its experiment, and Bootstrap
should not be ran until the next step.
Alternatively, a single model can also be trained at once by calling cxr_fairness/train.py
with the appropriate arguments, for example:
python -m cxr_fairness.train \
--algorithm DistMatch \
--distmatch_penalty_weight 5.0 \
--match_type mean \
--batch_size 32 \
--data_type balanced \
--dataset CXP \
--output_dir {output_dir} \
--protected_attr sex \
--task "No Finding" \
--val_fold 0
After all experiments have finished, run the notebooks/get_best_model_configs.ipynb
notebook to select the best models across hyperparameter settings.
Then, run the Bootstrap
experiment using cxr_fairness.sweep
as shown above, updating the path in experiments.py
appropriately.
We provide the following notebooks in the notebooks
folder to create figures shown in the paper:
agg_results_single_target.ipynb
: Creates two main result figures (performance metrics and comparison with Balanced ERM) for a task and dataset.adv_perf_graph.ipynb
: Creates the figure showing performance of group fairness methods as a function of the loss term weighting.We provide the radiologist labelled dataset at aux_data/MIMIC_CXR_Rad_Labels.xlsx
, and we analyze the dataset using notebooks/rad_labels.ipynb
.
We link the x-rays in MIMIC-CXR with hospital stay information in MIMIC-IV by querying MIMIC-IV through Google BigQuery in a Colab Notebook. The resulting data can then be downloaded, and is merged with the model predictions and analyzed using notebooks/proxy_label_graphs.ipynb
.
If you use this code in your research, please cite the following publication:
@article{zhang2022improving,
title={Improving the Fairness of Chest X-ray Classifiers},
author = {Zhang, Haoran and Dullerud, Natalie and Roth, Karsten and Oakden-Rayner, Lauren and Pfohl, Stephen Robert and Ghassemi, Marzyeh},
journal={arXiv preprint arXiv:2203.12609},
year={2022}
}