Multimodal Clinical Pseudo-notes for Emergency Department Prediction Tasks using Multiple Embedding Model for EHR (MEME)

Abstract

In this work, we introduce Multiple Embedding Model for EHR (MEME), an approach that views Electronic Health Records (EHR) as multimodal data. It uniquely represents tabular concepts like diagnoses and medications as structured natural language text using our "pseudo-notes" method. This approach allows us to effectively employ Large Language Models (LLMs) for individual EHR representation, proving beneficial in a variety of text-classification tasks. We demonstrate the effectiveness of MEME by applying it to diverse tasks within the Emergency Department across multiple hospital systems. Our findings show that MEME surpasses the performance of both single modality/embedding methods and traditional machine learning approaches, highlighting its effectiveness. Additionally, our tests on the model's generalizability reveal that training solely on the MIMIC-IV database does not guarantee effective application across different hospital institutions.

Huggingface Repository

Below is the tree structure of the repository, listing all the model files and their respective functions:

MEME-repository/
├── .gitattributes
├── README.md
├── MEME-disposition-final.pth
├── MEME-multitask-final.pth
├── MSEM-disposition.pth
├── MSEM-multitask.pth
├── arrival-disposition-final.pth
├── arrival-multitask-final.pth
├── codes-disposition-final.pth
├── codes-multitask-final.pth
├── medrecon-disposition-final.pth
├── medrecon-multitask-final.pth
├── pyxis-disposition-final.pth
├── pyxis-multitask-final.pth
├── triage-disposition-final.pth
├── triage-multitask-final.pth
├── vitals-disposition-final.pth
└── vitals-multitask-final.pth

Usage

The models are trained to perform specific tasks related to the emergency department using Multiple Embedding Model for EHR (MEME), Multimodal Single Embedding Model (MSEM), and modality specific single embedding models. They are designed to predict various outcomes and assist in multitask and disposition prediction tasks.

To use these models, load them into your PyTorch environment using the following example code:

import torch

# Example of loading the MEME disposition model
model = torch.load('MEME-disposition-final.pth')

# Your code to use the model goes here

Contributing

If you wish to contribute to this repository, please fork it, make your changes, and submit a pull request.

For any questions or issues, please open an issue on this repository or reach out to simonlee711@g.ucla.edu

Thank you for your interest in artificial intelligence within Healthcare.