kithminrw / ecg-recon

ECG reconstruction

Twelve-Lead ECG Reconstruction from Single-Lead Signals Using Generative Adversarial Networks

Repository for Twelve-Lead ECG Reconstruction from Single-Lead Signals Using Generative Adversarial Networks published in MICCAI 2023

Jinho Joo¹*, Gihun Joo¹*, Yeji Kim², Moo-Nyun Jin², Junbeom Park²**, Hyeonseung Im¹**

¹Kangwon National University, ²Ewha Womans University Medical Center

* denotes equal contribution, ** denotes corresponding authors

Our Goal

We propose a novel generative adversarial network that can faithfully reconstruct 12-lead ECG signals from single-lead signals. Our method can reconstruct 12-lead ECGs with CVD-related characteristics effectively. Thus, our method can be used to bridge commonly available wearable devices that can measure only Lead I and high-performance deep learning-based prediction models using 12-lead ECGs.

EKGAN Architecture

Experiments

We implemented not only EKGAN but also Pix2pix, CycleGAN, and CardioGAN with minor modifications so that they can be applied to ECG data. Additionally, 12-lead ECGs were generated by using both the validation and test sets, and their quality was evaluated by using a CVD prediction model, comparing the classification performance with the original 12-lead ECGs and the generated ones, and examined by three cardiologists.

Citation

@inproceedings{joo2023twelve,
  title={Twelve-Lead ECG Reconstruction from Single-Lead Signals Using Generative Adversarial Networks},
  author={Joo, Jinho and Joo, Gihun and Kim, Yeji and Jin, Moo-Nyun and Park, Junbeom and Im, Hyeonseung},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={184--194},
  year={2023},
  organization={Springer}
}

Q&A

If you have a question regarding the code, please email at jinho381 AT naver DOT com or joo9327 AT naver DOT com.