Algorithmic encoding of protected characteristics in chest X-ray disease detection models

This repository contains the code for the paper

B. Glocker, C. Jones, M. Bernhardt, S. Winzeck
Algorithmic encoding of protected characteristics in chest X-ray disease detection models
eBioMedicine. Volume 89, 104467, March 2023.

Dataset

The CheXpert imaging dataset together with the patient demographic information used in this work can be downloaded from https://stanfordmlgroup.github.io/competitions/chexpert/.

The MIMIC-CXR imaging dataset can be downloaded from https://physionet.org/content/mimic-cxr-jpg/2.0.0/ with the corresponding demographic information available from https://physionet.org/content/mimiciv/1.0/.

Code

For running the code, we recommend setting up a dedicated Python environment.

Setup Python environment using conda

Create and activate a Python 3 conda environment:

conda create -n chexploration python=3
conda activate chexploration

Install PyTorch using conda (for CUDA Toolkit 11.3):

conda install pytorch torchvision cudatoolkit=11.3 -c pytorch

Setup Python environment using virtualenv

Create and activate a Python 3 virtual environment:

virtualenv -p python3 <path_to_envs>/chexploration
source <path_to_envs>/chexploration/bin/activate

Install PyTorch using pip:

pip install torch torchvision

Install additional Python packages:

pip install matplotlib jupyter pandas seaborn pytorch-lightning scikit-learn scikit-image tensorboard tqdm openpyxl tabulate statsmodels

Requirements

The code has been tested on Windows 10 and Ubuntu 18.04/20.04 operating systems. The data analysis does not require any specific hardware and can be run on standard laptop computers. The training and testing of the disease detection models requires a high-end GPU workstation. For our experiments, we used a NVIDIA Titan X RTX 24 GB.

How to use

In order to replicate the results presented in the paper, please follow these steps:

1. Download the CheXpert dataset, copy the file train.csv to the datafiles/chexpert folder. Download the CheXpert demographics data, copy the file CHEXPERT DEMO.xlsx to the datafiles/chexpert folder.
2. Download the MIMIC-CXR dataset, copy the files mimic-cxr-2.0.0-metdata.csv and mimic-cxr-2.0.0-chexpert.csv to the datafiles/mimic folder. Download the MIMIC-IV demographics data, copy the files admissions.csv and patients.csv to the datafiles/mimic folder.
3. Run the notebooks chexpert.sample.ipynb and mimic.sample.ipynb to generate the study data.
4. Run the notebook chexpert.resample.ipynb to perform test-set resampling.

To replicate the results on CheXpert:

1. Adjust the variable img_data_dir to point to the CheXpert imaging data and run the following scripts:
   • Run the script chexpert.disease.py to train a disease detection model.
   • Run the script chexpert.sex.py to train a sex classification model.
   • Run the script chexpert.race.py to train a race classification model.
   • Run the script chexpert.multitask.py to train a multitask model.
2. Run the notebook chexpert.predictions.ipynb to evaluate all the prediction models.
3. Run the notebook chexpert.explorer.ipynb for the unsupervised exploration of feature representations.

Additionally, there are scripts chexpert.sex.split.py and chexpert.race.split.py to run SPLIT on the disease detection model. The default setting in all scripts is to train a DenseNet-121 using the training data from all patients. The results for models trained on subgroups only can be produced by changing the path to the data files (e.g., using chexpert.sample.train.white.csv and chexpert.sample.val.white.csv instead of chexpert.sample.train.csv and chexpert.sample.val.csv).

To replicate the results on MIMIC-CXR, adjust the above scripts accordingly to point to the MIMIC imaging data and its corresponding data files (e.g., mimic.sample.train.csv, mimic.sample.val.csv and mimic.sample.test.csv).

Note, the Python scripts also contain code for running the experiments using a ResNet-34 backbone which requires less GPU memory than DenseNet-121.

Trained models

All trained models, feature embeddings and output predictions can be found here. These can be used to directly reproduce the results presented in our paper using the notebooks chexpert.predictions.ipynb and chexpert.explorer.ipynb.

Expected outputs and runtimes

The scripts for disease detection, sex classification, and race classification will produce outputs in csv format that can be processed by the evaluation code in the Jupyer notebooks. The notebooks will produce figures and plots either in png or pdf format.

Training the models for disease detection, sex classification, and race classification will take about three hours each on a high-end GPU workstation. Running the data analysis code provided in the notebooks takes several minutes on a standard laptop computer.

Funding sources

This work is supported through funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (Grant Agreement No. 757173, Project MIRA, ERC-2017-STG) and by the UKRI London Medical Imaging & Artificial Intelligence Centre for Value Based Healthcare.

License

This project is licensed under the Apache License 2.0.