Original version of this repository can be found from here. It contains the Pytorch implementation of the ResNet model by Between_a_ROC_and_a_heart_place. The model was designed for the PhysioNet/Computing in Cardiology Challenge 2020. The related paper was accepted by the CinC2020 and titled "Adaptive Lead Weighted ResNet Trained With Different Duration Signals for Classifying 12-lead ECGs".

This version is refactored for more general analysis.

Install the required pip packages from requirements.txt file using the following command:

pip install -r requirements.txt

Recommended Python version 3.10.4 (tested with Python 3.10.4).

Check out the notebook Introduction to data handling in /notebooks/ for more information on downloading, preprocessing and splitting data.

If you want to preprocess data, you can do it with the preprocess_data.py script. This is not mandatory for the use of the repository, but keep in mind that if some transforms (e.g. BandPassFilter) are used during the training phase, training might slow down significantly. To preprocess the data, use the following command

python preprocess_data.py

Consider checking the configs directory for yaml configurations:

• Yaml files in the training directory are used to train a model
• Yaml files in the predicting directory are used to test and evaluate a model

Two notebooks are available for creating training and testing yaml files based on the data splitting performed with the create_data_csvs.py script: Yaml files of database-wise split for training and testing and Yaml files of stratified split for training and testing. Be sure to perform the data splitting first.

1. To split the data for the model to use in training and testing, you'll need the following command

python create_data_csvs.py

where create_data_csvs.py splits the data using either stratified split or database-wise split. On stratified run, create_data_csvs.py uses the implementation of MultilabelStratifiedShuffleSplit from iterative-stratification package. It makes csv files of the data splits which consists of a training set and a validation set. These csv files are later used in the training phase of the model, and have the columns path (path for ECG recording in .mat format), age , gender and all the diagnoses in SNOMED CT codes used as labels in the classification. Csv files of test data are also created. Database-wise split uses the structure of the directory where the data is loaded from.

The main structure of csv files are as follows:

Note! There are attributes to be considered before running the script. Check the notebook Introduction to data handling for further instructions.

2. To train a model, you'll need to use either a yaml file or a directory as an argument and use one of the following commands

python train_model.py train_smoke.yaml
python train_model.py train_stratified_smoke

where train_data.yaml consists of needed arguments for the training in a yaml format, and train_multiple_smoke is a directory containing several yaml files. When using multiple yaml files at the same time, each yaml file is loaded and run separately. More detailed information about training is available in the notebook Introduction to training models.

3. To test and evaluate a trained model, you'll need one of the following commands

python run_model.py predict_smoke.yaml
python run_model.py predict_stratified_smoke

where predict_smoke.yaml consists of needed arguments for the prediction phase in a yaml format, and predict_multiple_smoke is a directory containing several yaml files. When using multiple yaml files at the same time, each yaml file is loaded and run separately. More detailed information about prediction and evaluation is available in the notebook Introduction to testing and evaluating models.

.
├── configs                      
│   ├── data_splitting           # Yaml files considering a database-wise split and a stratified split   
│   ├── predicting               # Yaml files considering the prediction and evaluation phase
│   └── training                 # Yaml files considering the training phase
│   
├── data
│   ├── smoke_data               # Samples from the Physionet 2021 Challenge data as well as
|   |                              Shandong Provincial Hospital data for smoke testing
│   └── split_csvs               # Csv files of ECGs, either database-wise or stratified splitted
│
├── notebooks                    # Jupyter notebooks for data exploration and 
│                                  information about the use of the repository
├── src        
│   ├── dataloader 
│   │   ├── __init__.py
│   │   ├── dataset.py           # Script for custom DataLoader for ECG data
│   │   ├── dataset_utils.py     # Script for preprocessing ECG data
│   │   └── transforms.py        # Script for tranforms
│   │
│   └── modeling 
│       ├── models               # All model architectures
│       │   └── seresnet18.py    # PyTorch implementation of the SE-ResNet18 model
│       ├──__init__.py
│       ├── metrics.py           # Script for evaluation metrics
│       ├── predict_utils.py     # Script for making predictions with a trained model
│       └── train_utils.py       # Setting up optimizer, loss, model, evaluation metrics
│                                  and the training loop
│
├── .gitignore
├── label_mapping.py             # Script to convert other diagnostic codes to SNOMED CT Codes
├── LICENSE
├── LICENSE.txt
├── __init__.py
├── create_data_csvs.py          # Script to perform database-wise data split or split by
│                                  the cross-validatior ´Multilabel Stratified ShuffleSplit´ 
├── preprocess_data.py           # Script for preprocessing data
├── README.md
├── requirements.txt             # The requirements needed to run the repository
├── run_model.py                # Script to test and evaluate a trained model
├── train_model.py               # Script to train a model
└── utils.py                     # Script for yaml configuration