MLD3/disparate_censorship

Disparate Censorship: A Mechanism for Model Performance Gaps in Clinical Machine Learning

Official code repository for the paper "Disparate Censorship: A Mechanism for Model Performance Gaps in Clinical Machine Learning" (MLHC 2022).

Experiments

All scripts for experiments are in the scripts/ directory.

• Setting 2: Run the scripts under scripts/ in this directory (i.e. bash scripts/setting2_full.csv).
  • scripts/setting2_full.csv: Performance gap results for models at varying testing thresholds. See simulation_results.ipynb to access/plot results.
  • scripts/setting2_variance.csv, scripts/setting2_mean_distance.csv: Supplementary results under different distributional parameters at a fixed testing threshold. See simulation_results.ipynb to access/plot results.
• Setting 3: Run the Python script script/conditional_shift_experiments.py. See simulation_results.ipynb to access/plot results.

Raw results will be outputted into the results/ directory, where the results for each individual run (i.e. threshold/distribution parameter setting) and associated replications are stored in CSV files (e.g. results/SETTING2_BASE_5_5/*.csv). See simulation_results.ipynb for an example of how to read the raw results.

Figures

Plots can be reconstructed using each of the notebooks:

• simulation_examples.ipynb: 2D toy examples of simulated disparate censorship
• simulation_results.ipynb: Plots for performance gaps in Setting 2 and 3 of the paper
• mimic_results.ipynb: Plot (singular) and hypothesis test results for all MIMIC-IV v1.0 results

All figures will be outputted into the figures/ directory.

Citation

If you use this repository or our paper for your own research, we'd love for you to cite our work:

@inproceedings{chang2022disparate,
  title={Disparate Censorship: A Mechanism for Model Performance Gaps in Clinical Machine Learning},
  author={Chang, Trenton and Sjoding, Michael W and Wiens, Jenna},
  booktitle={Machine Learning for Healthcare Conference},
  volume={182},
  year={2022},
  organization={PMLR}
}