We use the MIMIC-III and MIMIC-IV datasets to benchmark our method.
The download of the data is available at PhysioNet. You need to complete a short course to obtain access as required by the data issuer. Once you download the data in tabular form, you can construct the graph using get_graph.py.
The processed data (in pkl formats) will be stored in the respective subdirectory under the data folder. You may call main.py to start a training. Exemplar training configurations are provided in ./config in yaml formats. Benchmarking is also avaialble in benchmark.py. Testing performance will be recorded after every epoch. We adopt wandb for results management and results will be uploaded to wandb online if switched on.
We implement our method based on the pyhealth package.
If you find our work useful, please cite us at
@article{chan2024multi,
title={Multi-task heterogeneous graph learning on electronic health records},
author={Chan, Tsai Hor and Yin, Guosheng and Bae, Kyongtae and Yu, Lequan},
journal={Neural Networks},
pages={106644},
year={2024},
publisher={Elsevier}
}