Description

The Python implementation for the publication “Uncertainty-Aware Time-to-Event Prediction using Deep Kernel Accelerated Failure Time Models” on MLHC 2021 (Machine Learning for Healthcare 2021) using PyTorch and GPyTorch packages.

Project structure

.
├── data_utils.py
├── gp_layer.py
├── LICENSE
├── logging_conf.py
├── model_utils.py
├── plot_utils.py
├── pml_los.py
├── pml_pfs.py
├── pytorchtools.py
├── README.md
├── requirements.txt
├── run_exp_gp_los_metric.py
├── run_exp_gp_los.py
├── run_exp_gp_pfs_exact_metric.py
├── run_exp_gp_pfs_exact.py
├── run_exp_gp_pfs_svgp_metric.py
├── run_exp_gp_pfs_svgp.py
├── run_exp_los_dropout.py
├── run_exp_los_metric.py
├── run_exp_los.py
├── run_exp_pfs_dropout.py
├── run_exp_pfs_metric.py
└── run_exp_pfs.py

Usage

• The access to datasets for the prediction of Progression Free Survival (PFS) and Length-of-Stay (LoS) has to be applied before running the code.
  • THe prediction of PFS involves the PRAEGNANT Dataset.
  • The prediction of LoS involves the MIMIC dataset with the data pre-processing from Purushotham et al..
• All .py files should be able to run with python xxx.py after installing the packages specified in requirements.txt.
• The .py scripts prefixed with run_exp_ can be used to build models proposed in the paper.
  • Scripts with …_pfs_… are for the task of PFS prediction.
  • Scripts with …_los_… are for the task of LoS prediction.

Note

The code is published to ensure the reproducibility in the machine learning community. If you find the code helpful, please consider citing

@article{wu2021uncertainty,
  title={Uncertainty-Aware Time-to-Event Prediction using Deep Kernel Accelerated Failure Time Models},
  author={Wu, Zhiliang and Yang, Yinchong and Fasching, Peter A and Tresp, Volker},
  journal={arXiv preprint arXiv:2107.12250},
  year={2021}
}

License

The code has a MIT license, as found in the LICENSE file.