deepfake-ecg
To generate synthetic ECGs from Hugging face
from transformers import AutoModel
model = AutoModel.from_pretrained("deepsynthbody/deepfake_ecg", trust_remote_code=True)
out = model(num_samples=5)
Pulse2Pulse - development repo
If you want to train the model from scratch, please refere our development repository Pulse2Pulse.
Usage
The generator functions can generate DeepFake ECGs with 8-lead values [lead names from first coloum to eighth colum: 'I','II','V1','V2','V3','V4','V5','V6'] for 10s (5000 values per lead). These 8-leads format can be converted to 12-leads format using the following equations.
lead III value = (lead II value) - (lead I value)
lead aVR value = -0.5*(lead I value + lead II value)
lead aVL value = lead I value - 0.5 * lead II value
lead aVF value = lead II value - 0.5 * lead I value
Pre-generated DeepFake ECGs and corresponding MUSE reports are here: https://osf.io/6hved/ or (https://huggingface.co/datasets/deepsynthbody/deepfake_ecg)
- In this repository, there are two DeepFake datasets:
1. 150k dataset - Randomly generated 150k DeepFakeECGs
2. Filtered all normals dataset - Only "Normal" ECGs filtered using the MUSE analysis report
A real ECG vs a DeepFake ECG (from left to right):
A sample DeepFake ECG:
Contributing
Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.
Please make sure to update tests as appropriate.
Citation:
@article{thambawita2021deepfake,
title={DeepFake electrocardiograms using generative adversarial networks are the beginning of the end for privacy issues in medicine},
author={Thambawita, Vajira and Isaksen, Jonas L and Hicks, Steven A and Ghouse, Jonas and Ahlberg, Gustav and Linneberg, Allan and Grarup, Niels and Ellervik, Christina and Olesen, Morten Salling and Hansen, Torben and others},
journal={Scientific reports},
volume={11},
number={1},
pages={1--8},
year={2021},
publisher={Nature Publishing Group}
}
License
For more details:
Please contact: vajira@simula.no, michael@simula.no