MedM2T: A MultiModal Framework for Time-Aware Modeling with Electronic Health Record and Electrocardiogram Data

This repository contains the official implementation of MedM2T, as described in our paper:

MedM2T: A MultiModal Framework for Time-Aware Modeling with Electronic Health Record and Electrocardiogram Data
Yu-Chen Kuo and Yi-Ju Tseng
(submitted to IEEE Journal of Biomedical and Health Informatics)

Supplementary Material

The supplementary materials provide extended tables, figures, and detailed dataset descriptions.

• Supplementary Document
• Additional Supplementary Files

MedM2T Repository Structure

• src/: Implementation of MedM2T framework.
• dataset/: Dataset class definitions and preprocessing scripts for EHR and ECG data.

Data Availability

This project uses the MIMIC-IV database (version 2.2) and MIMIC-IV-ECG database (version 1.0) as the primary data source.

Due to the restricted-access policy, we cannot release any raw data within this repository.
Accessing MIMIC-IV requires completing the required training and obtaining approval for access through PhysioNet: PhysioNet Credentialing Process

Preprocessing Pipelines

To ensure reproducibility, we provide source code for preprocessing in the FirstICU folder:

• Task 2 (In-hospital Mortality Prediction)
• Task 3 (Length of Stay Prediction)

These scripts demonstrate the pipeline used to process the raw MIMIC-IV data.
Users with authorized access to MIMIC-IV can follow these pipelines to reproduce the datasets for training and evaluation.

Acknowledgment

The ResNet in this project is adapted from antonior92/automatic-ecg-diagnosis.
We have modified and extended the original implementation to fit the MedM2T framework.
If you use this code, please also cite the original repository/paper as indicated by the authors.