How to Leverage Multimodal EHR Data for Better Medical Predictions?

This repository contains the code of the paper: How to Leverage Multimodal EHR Data for Better Medical Predictions?

Installation

All the dependencies are in the requirements.txt. You can build the environment with the following command:

conda create --name --file requirements.txt

Data Download

The MIMIC-III data can be downloaded at: https://physionet.org/content/mimiciii/1.4/. This dataset is a restricted-access resource. To access the files, you must be a credentialed user and sign the data use agreement (DUA) for the project. Because of the DUA, we cannot provide the data directly.

The pre-trained parameters of ClinicalBERT can be downloaded at: https://github.com/kexinhuang12345/clinicalBERT.

Instructions

We provide the default hyperparameters in params.py, you can modify it to try different hyperparamters. To run the code in this folder, please follow the instructions below.

1. Download the MIMIC-III data.
2. Extract the features by the scripts provided at: https://github.com/MLD3/FIDDLE-experiments
3. Set the mimic_dir and data_dir in data_module.py to the path of the above data.
4. Set the task related information and run data_module.py to combine clinical notes with other data.
5. Run the model by: python run.py --task=task_name --model=model_name