🩻 MedMod: A Multimodal Benchmark for Clinical Prediction Tasks with Electronic Health Records and Chest X-Ray Scans

Table of contents

• Background
• Environment setup
• Model training
• Model evaluation
• Citation

Background

We follow the data extraction and linking pipeline of the two datasets MIMIC-IV and MIMIC-CXR based on the task definition (i.e., inhospital mortality prediction, clinical conditions, decompensation, length of stay, and radiology).

Environment setup

To run this repo, you must install and run the libraries in the below yml file.

conda env create -f environment.yml
conda activate medmod

Training and evaluation framework

Self-supervised pre-training scripts

For pre-training, there are three types of training scripts that have been setup:

• simclr_trainer.py
• vicreg_trainer.py
• align_trainer.sh

Self-supervised evaluation scripts

To evaluate the quality of the representations learned using pre-training, several scripts have been implemented:

• (task/train/script.sh) finetune.sh, finetune_cxr.sh, finetune_ehr.sh --> these scripts further fine-tune either the multi-modal architecture or the uni-modal branches.
• (task/train/script.sh) lineareval.sh, lineareval_cxr.sh, lineareval_ehr.sh --> these scripts tune a single layer and freeze the pre-trained encoders for either multi-modal or uni-modal predictions.
• (task/eval/script.sh) lineareval / fteval --> these scripts perform an evaluation run for either fine-tuned or linear classifiers.

All of the scripts above call run_gpu.py

Other useful scripts:

• eval_epoch.sh --> This calls epoch_evaluation.py and evaluates the quality of the representations in terms of AUROC using a linear classifier at each pre-training epoch. It is useful for selecting the epoch that yields the best AUROC on the validation set. It stores results in a csv file.

Citation

If you find MedMod useful for your research and applications, please cite using this BibTeX:

@misc{author2024medmod,
    title={MedMod: A Multimodal Benchmark for Clinical Prediction Tasks with Electronic Health Records and Chest X-Ray Scans},
    url={https://anonymous.4open.science/r/MedMod-9061/},
    author={Authors},
    month={February},
    year={2025}
}