Multimodal Patient Representation Learning with Missing Modalities and Labels (ICLR 2024)

This repository contains code for the ICLR'24 paper: Multimodal Patient Representation Learning with Missing Modalities and Labels.

Dependencies

python==3.8.18
torch==2.0.1

Repository Structure

• src/: Source code for MedLink
  • preprocess/: Scripts for data preprocessing
  • dataset/: Data, Dataset, Tokenizer, Vocabulary, and collate_fn
  • core/: Core implementation for the MUSE method
  • metrics.py: Metrics for model evaluation
  • helper.py: Helper class for model training, evaluation, and inference
  • utils.py: Utility functions

How to Reproduce

Follow these steps to reproduce the results:

1. Edit the path in src/utils.py to your local path.
2. Obtain the eICU and MIMIC-IV datasets and place it under {raw_data_path}.
3. Run the following notebooks under src/preprocess in the specified order to prepare the data:
   1. eICU:
      1. Run parse_eicu_remote.ipynb
      2. Run preprocess_eicu.py
      3. Run build_vocab_eicu.py
      4. Run data_split_eicu.py
   2. MIMIC-IV:
      1. Run parse_mimic4_remote.ipynb
      2. Run preprocess_mimic4.py
      3. Run build_vocab_mimic4.py
      4. Run data_split_mimic4.py
   3. Run get_code_embeddings.py
4. Execute run.py under src/core to train the model:

   python run.py \
   --dataset [mimic4/eicu] \
   --task [mortality/readmission] \
   --official_run
   

Citation

@inproceedings{
wu2024multimodal,
title={Multimodal Patient Representation Learning with Missing Modalities and Labels},
author={Zhenbang Wu and Anant Dadu and Nicholas Tustison and Brian Avants and Mike Nalls and Jimeng Sun and Faraz Faghri},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=Je5SHCKpPa}
}