This repository includes the code of the ECG-DualNet for ECG classification proposed in the paper Exploring Novel Algorithms for Atrial Fibrillation Detection by Driving Graduate Level Education in Medical Machine Learning (Physiological Measurement).
This work was done original as part of the competition "Wettbewerb künstliche Intelligenz in der Medizin" at TU Darmstadt (KIS*MED, Prof. Hoog Antink).
A report of the project is available here. Slides of the final presentation are available here. LaTeX code of both the report and the slides are also available.
If you find this research useful in your work, please cite our paper:
@article{Rohr2022,
title={{Exploring Novel Algorithms for Atrial Fibrillation Detection by Driving Graduate Level Education in Medical Machine Learning}},
author={Rohr, Maurice and Reich, Christoph and H{\"o}hl, Andreas and Lilienthal, Timm and Dege, Tizian and Plesinger, Filip and Bulkova, Veronika and Clifford, Gari D and Reyna, Matthew A and Antink, Christoph Hoog},
journal={{Physiological Measurement}},
year={2022},
publisher={IOP Publishing}
}
All dependencies can be installed by running the following commands:
git clone https://github.com/ChristophReich1996/ECG_Classification
cd ECG_Classification
pip install --no-deps -r requirements.txt -f https://download.pytorch.org/whl/torch_stable.html
cd ecg_classification/pade_activation_unit/cuda
python setup.py install
cd ../../../
The implementation was tested Gentoo Linux 5.10.7, Python 3.8.5, and CUDA 11.1. To perform training and validation a CUDA device is recommended! This is due to the PAU implementation, which provides an efficient CUDA implementation. Training and inference on the CPU is supported but not recommended. The functionality of this repository can not be guaranteed for other system configurations.
If only CUDA 11.0 is available the code can also be executed with PyTorch 1.7.1 and Torchaudio 0.7.2 see.
ECG-DualNet is implemented with PyTorch 1.8.1 and
Torchaudio 0.8.1. All required packages can be seen
in the requirements.txt
file. For the Pade Activation Unit the
implementation (cuda extension) from the authors were adopted [1].
This repository also offers a Dockerfile to install all dependencies.
To build the Docker image run:
docker build --tag ecg_classification .
To execute the container run (with all available GPUs):
docker run -it --gpus all --rm --name user_name ecg_classification
This Dockerfile does not use the PAU CUDA extension, which leads to a higher memory usage and an increased runtime.
A Dockerfile based on the Nvidia NGC container, supporting the PAU CUDA extension, is also available.
Three different training runs are always reported. Weights of the best-performing model are provided.
The following validation results for the custom training/validation split on the 2017 PhysioNet Challenge dataset [2] (without pre-training) were achieved.
Model | ACC | F1 | # Parameters | best |
---|---|---|---|---|
ECG-DualNet S (CNN + LSTM) | 0.8527; 0.8410; 0.8455 | 0.8049; 0.7923; 0.7799 | 1840210 | weights |
ECG-DualNet M (CNN + LSTM) | 0.8560; 0.8442; 0.8495 | 0.7938; 0.7955; 0.7928 | 4269618 | weights |
ECG-DualNet L (CNN + LSTM) | 0.8508; 0.8213; 0.8514 | 0.8097; 0.7515; 0.8038 | 6176498 | weights |
ECG-DualNet XL (CNN + LSTM) | 0.7702; 0.8612; 0.7866 | 0.6899; 0.8164; 0.7162 | 20683122 | weights |
ECG-DualNet++ S (AxAtt + Trans.) | 0.7323; 0.8174; 0.7912 | 0.6239; 0.7291; 0.7127 | 1785186 | weights |
ECG-DualNet++ M (AxAtt + Trans.) | 0.8226; 0.8259; 0.7938 | 0.7544; 0.7730; 0.6947 | 2589714 | weights |
ECG-DualNet++ L (AxAtt + Trans.) | 0.8449; 0.8442; 0.8396 | 0.7859; 0.7750; 0.7671 | 3717426 | weights |
ECG-DualNet++ XL (AxAtt + Trans.) | 0.8593; 0.8351; 0.8501 | 0.8051; 0.7799; 0.7851 | 8212658 | weights |
ECG-DualNet++ 130M (AxAtt + Trans.) | 0.8475; 0.8534; 0.8462 | 0.7878; 0.7963; 0.7740 | 127833010 | weights |
Note that for the weights of ECG-DualNet XL an external link is provided.
For training on the Icentia11k dataset [3] we achieved the following results:
Model | ACC | F1 | # Parameters | best | 20 epochs |
---|---|---|---|---|---|
ECG-DualNet XL (CNN + LSTM) | 0.8989 | 0.5135 | 20683122 | weights | weights |
ECG-DualNet++ XL (AxAtt + Trans.) | 0.8899 | 0.5017 | 8212658 | weights | weights |
If fine tuning the pretrained networks on the PhysioNet dataset [2] the following results were achieved:
Model | ACC | F1 | # Parameters | best |
---|---|---|---|---|
ECG-DualNet XL (CNN + LSTM) | 0.8455; 0.7664; 0.8468 | 0.7911; 0.5880; 0.8014 | 20683122 | weights |
ECG-DualNet++ XL (AxAtt + Trans.) | 0.8475; 0.8481; 0.8469 | 0.7828; 0.7817; 0.7899 | 8212658 | weights |
In the challange setting (pretrained weights used) the following results were achieved:
Model | ACC | F1 | # Parameters | best |
---|---|---|---|---|
ECG-DualNet XL (CNN + LSTM) | 0.8840; 0.8820; 0.8080 | 0.8549; 0.8449; 0.7360 | 20683122 | weights |
ECG-DualNet++ XL (AxAtt + Trans.) | 0.8720; 0.8800; 0.8680 | 0.8276; 0.8494; 0.8403 | 8212658 | weights |
In the two class challange setting (pretrained weights used) the following results were achieved:
Model | ACC | F1 | # Parameters | best |
---|---|---|---|---|
ECG-DualNet XL (CNN + LSTM) | 0.9800 | 0.9288 | 20683122 | weights |
ECG-DualNet++ XL (AxAtt + Trans.) | 0.9760 | 0.9146 | 8212658 | weights |
To reproduce the presented results simply run (a single GPU is needed):
sh run_experiments.sh
This script trains all models listed in the table above except ECG-DualNet++ 130M. During training all logs are saved in the experiments folder
(produced automatically). Most logs are stored in the
PyTorch tensor format .pt
and can be loaded by
loss:torch.Tensor = torch.load("loss.pt"")
. Network weights are stored as a
state dict and can be loaded
by state_dict:Dict[str, torch.Tensor] = torch.load("best_model.pt")
.
To train the biggest ECG-DualNet++ with 130M parameters run:
python -W ignore train.py --cuda_devices "0, 1, 2, 3" --epochs 100 --batch_size 24 --physio_net --dataset_path "data/training2017/" --network_config "ECGAttNet_130M" --data_parallel
Four GPUs with 16GB are recommended. Reducing the batch size is a possible workaround if limited GPU memory is available.
To reproduce the presented ablation studies run:
sh run_ablations.sh
To perform pretraining on the Icentia11k dataset [3] run:
python -W ignore train.py --cuda_devices "0" --batch_size 100 --dataset_path "/dataset/icentia11k" --icentia11k --network_config "ECGAttNet_XL" --epochs 20
Pretraining with a batch size of 100 requres a GPU with at least 32GB. If a batch size of 50 is utilized a 16GB GPU is needed. Batch size can only be set in steps of 50.
To train the pretrained models on the 2017 PhysioNet dataset [2] run:
python -W ignore train.py --cuda_devices "0" --epochs 100 --batch_size 24 --physio_net --dataset_path "data/training2017/" --network_config "ECGAttNet_XL" --load_network "experiments/20_05_2021__18_32_19ECGAttNet_XL_icentia11k_dataset/models/20.pt"
or
python -W ignore train.py --cuda_devices "0" --epochs 100 --batch_size 24 --physio_net --dataset_path "data/training2017/" --network_config "ECGCNN_XL" --load_network "experiments/21_05_2021__12_15_06ECGCNN_XL_icentia11k_dataset/models/20.pt"
The challange submission can be reproduced by setting the additional flag --challange
. For the two class challange
submission add the flag --two_classes
.
Pleas note that the dataset or model paths as well as the cuda devices might change for different systems!
To run custom training the train.py
script can be used. This script takes the following commands:
Argument | Default Value | Info |
---|---|---|
--cuda_devices |
"0" | String of cuda device indexes to be used. Indexes must be separated by a comma. |
--no_data_aug |
False | Binary flag. If set no data augmentation is utilized. |
--data_parallel |
False | Binary flag. If set data parallel is utilized. |
--epochs |
100 | Number of epochs to perform while training. |
--lr |
1e-03 | Learning rate to be employed. |
--physio_net |
False | Binary flag. Utilized PhysioNet dataset instead of default one. |
--batch_size |
24 | Number of epochs to perform while training. |
--dataset_path |
False | Path to dataset. |
--network_config |
"ECGCNN_M" | Type of network configuration to be utilized. |
--load_network |
None | If set given network (state dict) is loaded. |
--no_signal_encoder |
False | Binary flag. If set no signal encoder is utilized. |
--no_spectrogram_encoder |
False | Binary flag. If set no spectrogram encoder is utilized. |
--icentia11k |
False | Binary flag. If set icentia11k dataset is utilized. |
--challange |
False | Binary flag. If set challange split is utilized. |
--two_classes |
False | Binary flag. If set two classes are utilized. Can only used with PhysioNet dataset and challange flag. |
All network hyperparameters can be found and adjusted in the config.py
file.
The following files for the challange submission are taken form the 18-ha-2010-pj
repo by Maurice Rohr and Prof. Hoog Antink.
wettbewerb.py
, predict_pretrained.py
, predict_trained.py
,
and score.py
.
Please not that the weights linked in the results have to be downloaded and put in the correct directories. For detailed
information pleas have a look at the predict.py
file. Additionally, the publicly available PhysioNet [2]
samples have to be downloaded for training.
The cleaned data of the challenge (also coming from [2]) as well as the publicly available PhysioNet [2] samples can be downloaded here.
The Icentia11k dataset [3] used for pretraining can be downloaded here.
For compiling the report please use pdfLaTeX. To compile the presentation slides LuaLaTeX is requred.
[1] @article{Molina2019,
title={Pad\'{e} Activation Units: End-to-end Learning of Flexible Activation Functions in Deep Networks},
author={Molina, Alejandro and Schramowski, Patrick and Kersting, Kristian},
journal={preprint arXiv:1907.06732},
year={2019}
}
[2] @inproceedings{Clifford2017,
title={{AF Classification from a Short Single Lead ECG Recording: the
PhysioNet/Computing in Cardiology Challenge 2017}},
author={Clifford, Gari D and Liu, Chengyu and Moody, Benjamin and Li-wei, H Lehman and Silva, Ikaro and Li, Qiao and Johnson, AE and Mark, Roger G},
booktitle={2017 Computing in Cardiology (CinC)},
pages={1--4},
year={2017},
organization={IEEE}
}
[3] @article{Tan2019,
title={Icentia11k: An unsupervised representation learning dataset for arrhythmia subtype discovery},
author={Tan, Shawn and Androz, Guillaume and Chamseddine, Ahmad and Fecteau, Pierre and Courville, Aaron and Bengio, Yoshua and Cohen, Joseph Paul},
journal={preprint arXiv:1910.09570},
year={2019}
}