advmask_ecg

Pretraining ECG Data with Adversarial Masking Improves Model Generalizability for Data-Scarce Tasks @ ML4H 2022

We adapt parts of our code from the following sources:

1. Self-Supervised Pre-Training of Networks with CLOCS (https://github.com/danikiyasseh/CLOCS)
2. Adversarial Masking for Self-Supervised Learning (https://github.com/YugeTen/adios)
3. Lead-agnostic Self-supervised Learning for Local and Global Representations of Electrocardiogram (https://github.com/Jwoo5/fairseq-signals)
4. In-depth Benchmarking of Deep Neural Network Architectures for ECG Diagnosis (https://github.com/seitalab/dnn_ecg_comparison)
5. torch_ecg (https://github.com/DeepPSP/torch_ecg/)

Datasets

CinC2020 (Physionet/Computing in Cardiology 2020)

Download WFDB zipped files (https://moody-challenge.physionet.org/2020/) to data/cinc2020/raw/ and unzip.

Generate train/val split by running python data/cinc2020/save_splits.sh

CinC2021 (Physionet/Computing in Cardiology 2021)

Download WFDB zipped files (https://moody-challenge.physionet.org/2021/) to data/cinc2021/raw/ and unzip. Merge WFDB_ChapmanShaoxing and WFDB_Ningbo to a folder named WFDB_ShaoxingUniv.

Generate train/val split by running python data/cinc2021/save_splits.sh

Chapman-Shaoxing

Download dataset (https://figshare.com/collections/ChapmanECG/4560497/2) to data/chapman/raw/

Generate train/val/test split by running python data/chapman/save_splits.sh

Run Example

Install requirements pip install requirements.txt. You will need a GPU for training with Pytorch-Lightning.

To run a single pretraining trial, run bash examples/bash_example.sh

To run a sweep over hyperparameters via slurm, run python examples/slurm_sweep_example.py