hawkiyc / WaveGAN_for_12_Leads_ECG_Signals

Notifications Before Continuing

For those who came here from 'Paper With Code' Website, this implementation is not exactly the same with the paper 'A Generative Adversarial Approach To ECG Synthesis And Denoising'. Here is something you shall be aware of:

1. This repository is only WaveGAN part, the rest is still in progress.
2. For computing intensity issue, I am using CODE-15% Dataset, NOT PTB-XL.
3. The hyperparameters are slightly different to the original article. You shall check and modify relevant hyperparameters in following files, "params.py, FetchData.py and WaveGAN.py", if you want fully reproduce the results of original paper.

WaveGAN for 12 Leads ECG Signals

12 Leads ECG is essential for identifying arrhythmia and other cardio malfunctions. However, it is always tricky to collect large amount of ECG data. With Deep Learning, generated ECG data can solve the problems of lacking data. GAN shows high performance in CV area for all candidates of generating model. However, due to the nature of time series data, it is harder to generate long sequence than big picture before WaveGAN architecture. Here by, I bulid a WaveGAN model for create artificial 12 Leads ECG data.

Here are results for this WaveGAN model

Dataset Description

I built this model with Code-15% dataset. You can find the description and original data from following link. https://zenodo.org/record/4916206

You can also download cleaned data for this particular project via following Link: https://1drv.ms/u/s!ArQCikHAsFj6oPJoQY7h9rIggjoaug?e=YjNprR

Reference:

1. https://github.com/mostafaelaraby/wavegan-pytorch

2. Adversarial Audio Synthesis: https://arxiv.org/abs/1802.04208

3. A Generative Adversarial Approach To ECG Synthesis And Denoising: https://www.researchgate.net/publication/344159398_A_Generative_Adversarial_Approach_To_ECG_Synthesis_And_Denoising