Skip to content

paidamoyo/counterfactual_survival_analysis

Repository files navigation

Enabling Counterfactual Survival Analysis with Balanced Representations (ACM CHIL 2021)

This repository contains the Pytorch code to replicate experiments in our paper Enabling Counterfactual Survival Analysis with Balanced Representations accepted at ACM Conference on Health, Inference, and Learning (ACM CHIL) 2021:

@inproceedings{chapfuwa2021enabling, 
  title={Enabling Counterfactual Survival Analysis with Balanced Representations},
  author={Chapfuwa, Paidamoyo and Assaad, Serge and Zeng, Shuxi and Pencina, Michael J and Carin, Lawrence and Henao, Ricardo},
  booktitle={ACM Conference on Health, Inference, and Learning},
  year={2021}
}

Model

Model

Data

  • ACTG: A longitudinal RCT study comparing monotherapy with Zidovudine or Didanosine with combination therapy in HIV patients
  • Framingham: A subset (Framingham Offspring) of the longitudinal study of heart disease dataset, for predicting the effects of statins on survival time
  • See actg_synthetic.ipynb to modify the generated ACTG-Synthetic data

Prerequisites

The code is implemented with the following dependencies:

pip install -r requirements.txt

Model Training

  • To train all models (CSA, CSA-INFO, AFT, AFT-Weibull, SR) run param_search.py
 python param_search.py

Metrics and Visualizations

Once the networks are trained and the results are saved, we extract the following key results:

Acknowledgments

This work leverages the calibration framework from SFM and the accuracy objective from DATE. Contact Paidamoyo for issues relevant to this project.