Multi-Channel Masked Autoencoder and Comprehensive Evaluations for Reconstructing 12-Lead ECG from Arbitrary Single-Lead ECG

Accepted by KDD workshop-AIDSH 224

First author: Jiarong Chen

Updates

Paper (Waiting for its publication, which included detailed information)
The application algorithm
The training algorithm
The testing algorithm
Pretrained weights
The samples for testing

Contributions

Reconstructing 12-Lead ECG from Arbitrary Single-Lead ECG
Comprehensive Evaluations, including signal-level, feature-level, and diagnostic-level

Saved models

An example model in MCMA could be seen in the Generator file
All the trained models: https://drive.google.com/drive/folders/1m57dz-FhcQCGNoZ2wxA_sUoHgrrGRHIn?usp=sharing

You can follow this repo for the newest information.

It is the open-source code for MCMA, which could reconstruct 12-lead ECG with arbitrary single-lead ECG. Before running, you should load your ECG signals, and the amplitude unit should be mv!!! If not, you should adjust it in advance. Before running, you should resample this signal as 500Hz. Before running, you can reshape it into (N,1024,1). Additionallu, the lead index should be provided. If not, you can try to find it by locating the maximum PCC. The lead index classifcaitionn accuracy in the internal testing dataset is 97.43%.

Citation

If you find this project is useful, please cite Multi-Channel Masked Autoencoder and Comprehensive Evaluations for Reconstructing 12-Lead ECG from Arbitrary Single-Lead ECG

@inproceedings{
chen2024multichannel,
title={Multi-Channel Masked Autoencoder and Comprehensive Evaluations for Reconstructing 12-Lead {ECG} from Arbitrary Single-Lead {ECG}},
author={Jiarong chen and Wanqing Wu and Shenda Hong},
booktitle={Artificial Intelligence and Data Science for Healthcare: Bridging Data-Centric AI and People-Centric Healthcare},
year={2024},
url={https://openreview.net/forum?id=lIX6BKDPJW}
}

@misc{chen2024multichannelmaskedautoencodercomprehensive,
      title={Multi-Channel Masked Autoencoder and Comprehensive Evaluations for Reconstructing 12-Lead ECG from Arbitrary Single-Lead ECG}, 
      author={Jiarong Chen and Wanqing Wu and Tong Liu and Shenda Hong},
      year={2024},
      eprint={2407.11481},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2407.11481}, 
}

Files

It includes demo.py, the trained model, and sample data.

Environment

conda env create -f environment.yml

Future work

I have tried different setting, but more efforts in model designing are necessary for this task.
High quality ECG, although this study based on the public dataset, the signal quality influence its evaluation and reconstruction.

Acknowledgements

Contacting me at chenjr356@gmail.com, chenjr56@mail2.sysu.edu.cn, jiarong.chen@sjtu.edu.cn

CHENJIAR3/MCMA