This repository contains all code and models relevant to ElectroCardioGuard (https://www.sciencedirect.com/science/article/pii/S0950705123007645, https://arxiv.org/abs/2306.06196).
Due to the large size of the IKEM dataset, we opted to publish it on Zenodo (https://zenodo.org/record/8393007).
pip install -r requirements.txt
To recreate experiments, please setup a MLflow tracking server according to instructions in dgn_mlflow_logger
.
Scripts for converting datasets to our HDF5 format are located in dataset_conversion_scripts
directory. In order to download all datasets and apply our pre-processing/compression to them, call
python download_datasets.py
The script will take a long time to fully complete (4-7 hours). To verify the process completed successfully, you can run python dataset_stats.py
and compare the numbers with this table:
Title | № ECGs | № patients | Size |
---|---|---|---|
PTB | 549 | 290 | 69MB |
PTB-XL | 21,799 | 18,869 | 2.1GB |
CODE-15% | 345,106 | 233,479 | 22GB |
In order to save disk space, we discard redundant leads (III, aVF, aVR, aVL). Specifically, the remaining leads V1-6, I, and II are stored in this order respectively from indices 0 to 7. We also quantize voltages to a 16-bit scale with 4.88 μV per bit. Tracings are stored in HDF5 files with a single tracing per chunk for fast random access.
results
contains full result tables, whose shortened and compact versions are published in our paper.
To run a single instance of grid search (model configuration), run python pt_grid_search_instance.py
with corresponding arguments.
To evaluate our model in the gallery/probe matching task or overseer simulation task, run pt_evaluate_as_classifier.py
with corresponding arguments. You can run gallery_probe_all.sh
and overseer_simulation_dev/test.sh
to directly reproduce our results, or update the model path parameter to evaluate a different model.
Our model is built on top of CDIL-CNN (models/pt_cdil_cnn.py
). The original implementation can be found here: https://github.com/LeiCheng-no/CDIL-CNN.
If you've found our work useful, please cite our publication:
@article{sejak2023electrocardioguard,
title={ElectroCardioGuard: Preventing Patient Misidentification in Electrocardiogram Databases through Neural Networks},
author={Sej{\'a}k, Michal and Sido, Jakub and {\v{Z}}ahour, David},
journal={arXiv preprint arXiv:2306.06196},
year={2023}
}