iTransplant: individualized TRANSparent Policy Learning for orgAN Transplantation (NeurIPS2021)

Source code for the iTransplant framework proposed in paper "Closing the loop in medical decision support by understanding clinical decision-making: A case study on organ transplantation". ITransplant is a novel data-driven framework to learn interpretable organ offer acceptance policies directly from clinical data. It learns a patient-wise parametrization of the expert clinician policy that accounts for the differences between patients, a crucial but often overlooked factor in organ transplantation. We conducted several investigative experiments with real-world liver transplantation data from the Organ Procurement and Transplantation Network (OPTN), covering 190,525 organ offers. The results show that iTransplant can be used to probe clinical decision-making practices in a number of ways. More specifically, our investigations allow us to: (1) identify important match criteria for organ offer acceptance; (2) discover patient-wise organ preferences of clinicians via automatic patient stratification in a latent representation space; and (3) examine the transplantation practice variations across transplant centers.

Installation & Environment Setup

To run the experiments locally, directly clone this repository via the following command.

git clone git@github.com:yvchao/iTransplant.git

We recommand creating the Python environment with pyenv (version 1.2.23) as follows.

pyenv install 3.8.7

Python of version 3.8.7 is preferred for best compatibility and reproducibility.

Make sure to install necessary dependencies with the provided requirements.txt before conducting any experiment with iTransplant.

pip install -r requirements.txt

Data Preparation

The liver transplantation data from OPTN is used for experiments in the paper. Please refer to the README.md file under the OPTN_data directory for instructions on how to obtain the dataset. The considered feature variables are listed in the selected_features.py file under the same path. Note that the liver transplantation data, waiting list history data and match run data (organ cllocation and associated decisions on organ offers) are required. The processed data need to be stored in HDF5 format with speficied keys. Please refer to src/data_loading/data_utils.py for details. Please put all processed data in the data folder to run experiments.

Main Results

The main results in the paper can be obtained by executing the following notebooks:

hyperparameter_selection.ipynb
benchmark.ipynb
investigative experiments.ipynb

Configurations used for hyperparameter selection and benchmark can be found in hyperparameter_selection_configures.py and benchmark_configures.py, respectively.

Additional benchmark results in the Appendix of the paper can be obtained by executing the following notebooks:

additional_hyperparameter_selection.ipynb
additional_benchmark.ipynb

Citation

If you find the software useful, please consider citing the following paper:

@inproceedings{iTransplant2021,
  title={Closing the loop in medical decision support by understanding clinical decision-making: A case study on organ transplantation},
  author={Qin, Yuchao and Imrie, Fergus and Hüyük, Alihan and Jarrett, Daniel and Edward Gimson, Alexander and van der Schaar, Mihaela},
  booktitle={Advances in neural information processing systems},
  year={2021}
}

Additional Notes

The experiments in the paper are conducted on a custom limited dataset (covering organ offers from January 1, 2003 to December 4, 2020) from OPTN.

There could be minor variations in the experiment results due to differences in:

Python versions
Versions of third-party dependencies.
Liver transplant data from OPTN

The iTransplant framework currently does not support execution on CUDA decives.

License

This software is released under the GPLv3 license.

yvchao / iTransplant