GAT-ETM

Graph ATtention-Embedded Topic Model (GAT-ETM) is an end-to-end graph-based embedded topic model that jointly learns a knowledge graph of medical codes (ICD disease and ATC drug codes) and patients EHRs. Our model is able to learn multimodal (disease and drug) topics based on co-occurence patterns from EHR and semantic similarity from knowledge graph. Our idea is to leverage existing knowledge to aid EHR modeling especially for rare medical codes. Notice that this framework is general and the ICD-ATC knowledge graph can be replaced by other knowledge graphs of the medical codes utilized in EHR data. This repo is the code used in a journal paper "Modeling electronic health record data using an end-to-end knowledge-graph-informed topic model". Please refer to the paper (https://www.nature.com/articles/s41598-022-22956-w) for details.

Overview of GAT-ETM

a. probabilistic graphic model (PGM) b. knowledge graph (KG) construction and augmentation. c. neural topic model (NTM) framework.

Preparing Data:

EHR data

Pre-trained code embeddings

• Metadata file (.txt) format is as follows:

    code1 code2 ...
    num_code1 num_code2 ...
    train_init_emb_flag
    init_emb_path

  where for the first two rows, each column describes type name and token vocabulary size of one type of codes, and the followed two rows are fix initilized embedding or not (1 or 0) and directory path of initialized embeddings file. We provide a sample metadata file.

Training model

    python3 main_getm.py --data_path data/your_data --save_path results/your_result --meta_file metadata_icd61210_atc82020_fine --mode train --gpu_device 0 --num_topics 100 --tq --epoch EPOCH --lr 0.01 

Notice: the metafile should be put in data_path directory

Loading Trained Model and Evalutating

    python3 main_getm.py --data_path data/your_data --save_path results/your_result --meta_file metadata_icd61210_atc82020_fine --mode eval -load_from results/your_trained_mdl --gpu_device 0 --num_topics 100 --tq --epoch EPOCH --lr 0.01  

Citation:

Please cite the following paper if you use this source in your work.

@article{zou2022modeling,
  title={Modeling electronic health record data using an end-to-end knowledge-graph-informed topic model},
  author={Zou, Yuesong and Pesaranghader, Ahmad and Song, Ziyang and Verma, Aman and Buckeridge, David L and Li, Yue},
  journal={Scientific Reports},
  volume={12},
  number={1},
  pages={17868},
  year={2022},
  publisher={Nature Publishing Group UK London}
}

For any clarification, comments, or suggestions please create an issue or contact Dr. Yue Li and Yuesong Zou.