Skip to content

Latest commit

 

History

History
38 lines (26 loc) · 2.12 KB

README.md

File metadata and controls

38 lines (26 loc) · 2.12 KB
Raw

MedML

Source code, knowledge graphs and feature lists for the paper MedML: Fusing Medical Knowledge and Machine Learning Models for Early Pediatric COVID-19 Hospitalization and Severity Prediction.

Requirements

  • Install python, pytorch. We use Python 3.6.15, Pytorch 1.7.1.
  • If you plan to use GPU computation, install CUDA

Data preparation

  • x_train/x_val/x_test are feature matrices to train the GNN model. Each matrix should has shape [N, F_g], where N denotes the number of train/val/test patients, and F_g denotes the feature dimension.
  • demo_train/demo_val/demo_test are additional feature matrices. These matrices stores the demographical data, which are not part of the knowledge graph but still will be used to train the model. Each matrix should has shape [N, F_d], where N denotes the number of train/val/test patients, and F_d denotes the feature dimension.
  • y_train/y_val/y_test are labels vectors. Each vector's length is N, where N denotes the number of train/val/test patients.
  • edge_pair.pkl stores the knowledge graph structure. It is a list of tuples. Each tuple (i,j) stores the edge from node i to j in the knowledge graph. Node index i and j should be corresponded to the feature index in the x matrices.

The data should be stored in the data folder. We provide a few dummy data for reference.

Training MedML

The training functions and hyperparameters are in the train_medml.py.

Citation

@article{gao2022medml,
  title={MedML: Fusing Medical Knowledge and Machine Learning Models for Early Pediatric COVID-19 Hospitalization and Severity Prediction},
  author={Gao, Junyi and Yang, Chaoqi and Heintz, Joerg and Barrows, Scott and Albers, Elise and Stapel, Mary and Warfield, Sara and Cross, Adam and Sun, Jimeng and others},
  journal={Iscience},
  pages={104970},
  year={2022},
  publisher={Elsevier}
}

License

Please refer to License.txt or License.docx for research usage.