Welcome to BioDG!

BioDG is a PyTorch suite containing benchmark datasets and algorithms for domain generalization, as introduced in Towards Domain Generalization for ECG and EEG Classification: Algorithms and Benchmarks.

The available algorithms were based on DomainBed and modified for 1D biosignal classification.

Available algorithms

The currently available ECG algorithms and currently available EEG algorithms are:

• Baseline - Empirical Risk Minimization (ERM, Vapnik, 1998)
• Invariant Risk Minimization (IRM, Arjovsky et al., 2019)
• Maximum Mean Discrepancy (MMD, Li et al., 2018)
• Deep CORAL (CORAL, Sun and Saenko, 2016)
• Representation Self-Challenging (RSC, Huang et al., 2020)

Available datasets

The currently available datasets are:

ECG

The datasets for the ECG Domain Generalization setup were taken from the 2020 PhysioNet Challenge and are the:

• China Physiological Signal Challenge 2018 (CPSC and CPSC Extra)
• PTB and PTB-XL Diagnostic ECG Database (PTB and PTB-XL)
• St Petersburg INCART 12-lead Arrhythmia Database (INCART)
• Georgia 12-Lead ECG Challenge Database (G12EC

To download the above datasets run the following in a terminal:

# Download all files
wget -r -N -c -np -R "index.html*" https://physionet.org/files/challenge-2020/1.0.2/training/

# INCART Annotations -- Extract data and rename folder to 'annotations'
wget -r -N -c -np -R "index.html*" https://physionet.org/files/incartdb/1.0.0/training/

mv physionet.org/files/challenge-2020/1.0.2/training/ .
mv physionet.org/files/challenge-2020/1.0.0/training/ annotations/

After downloading the datasets, please also download the following files:

• PhysioNet 2020 scored classes csv
• PhysioNet 2020 unscored classes csv

After extracting all datasets, the ECG data directory should follow the below tree structure:

├── annotations               
├── cpsc_2018
├── cpsc_2018_extra
├── georgia
├── ptb
├── ptb-xl
├── st_petersburg_incart
├── dx_mapping_scored.csv
└── dx_mapping_unscored.csv

EEG

The datasets for the EEG Domain Generalization are provided by the BCMI laboratory of the Shanghai Jiao Tong University, and are the following:

• SEED
• SEED-FRA
• SEED-GER

The datasets are available for research purposes, after applying here.

Quick start

ECG

1. First set the following variables in the bioconfig.py file:

• hostname --> Create a block with your hostname to set data paths
• scripts_root --> Root of the code package
• _root_ecg_path --> Root path of the ecg data

2. Convert the .mat ECG signal files to the appropriate format for the PyTorch DataLoader by running:

python3 ECG/convert_to_pickles.py --hostname user --outpath 'path to directory where converted data will be stored'

3. Set the following variables in the ECG bioconfig.py file:

• pickle_data_dir --> should be same path as the above output path
• ecg_results_dir --> experiment results path

4. Train a DG model:

python3 experiments/ecg_dg_train.py\
        --network RSC --algorithm RSC\
        --c "Flags are mentioned in the experiment file"

5. Train our proposed model:

python3 experiments/ecg.py --model biodg_resnet18 --epochs 30 --optim adam --batch_size 128

EEG

1. After downloading the datasets, run the following script to split and convert the EEG DE features to the appropriate format for the PyTorch DataLoader

python3 EEG/split_de_features.py\
        --eeg_de_features_path 'path to 1s_de_feature folder in EEG dataset'
        --split_data_path 'output path for split data'
        --dataset 'one between china, fra or ger'

2. Set the following variables in the EEG config.py file:

• pickle_data_dir --> should be the same as the split_data_path above
• eeg_results_dir --> experiment results path

3. Train a model:

python3 experiments/eeg_dg_train.py\
        --network ERM --algorithm ERM\
        --c "Flags are mentioned in the experiment file"

4. Train our proposed model:

python3 experiments/eeg.py --epochs 20 --optim adam --batch_size 128

Cite Us

If you use the above code for your research please cite our paper, which as of the 22nd of June 2023 has been accepted at IEEE TETCI:

@ARTICLE{10233054,
  author={Ballas, Aristotelis and Diou, Christos},
  journal={IEEE Transactions on Emerging Topics in Computational Intelligence}, 
  title={Towards Domain Generalization for ECG and EEG Classification: Algorithms and Benchmarks}, 
  year={2024},
  volume={8},
  number={1},
  pages={44-54},
  keywords={Brain modeling;Biological system modeling;Electrocardiography;Data models;Electroencephalography;Adaptation models;Feature extraction;Biosignal classification;deep learning;domain generalization;1D signal classification;electrocardiogram (ECG) classification;electroencephalogram (EEG) classification},
  doi={10.1109/TETCI.2023.3306253}}

If you are interested in other 1D biosignals, such as Phonocardiograms (PCG), check out our paper "Which Augmentation Should I Use? An Empirical Investigation of Augmentations for Self-Supervised Phonocardiogram Representation Learning", accepted in IEEE Access, here:

https://github.com/aristotelisballas/listen2yourheart/tree/main

@ARTICLE{10804781,
  author={Ballas, Aristotelis and Papapanagiotou, Vasileios and Diou, Christos},
  journal={IEEE Access}, 
  title={Which Augmentation Should I Use? An Empirical Investigation of Augmentations for Self-Supervised Phonocardiogram Representation Learning}, 
  year={2024},
  volume={12},
  number={},
  pages={193459-193472},
  keywords={Brain modeling;Phonocardiography;Biological system modeling;Training;Data models;Electrocardiography;Electroencephalography;Feature extraction;Robustness;Representation learning;Contrastive learning;deep learning;OOD representation learning;phonocardiogram classification;self-supervised learning},
  doi={10.1109/ACCESS.2024.3519297}}

License

This source code is released under the MIT license.