ECGAN is a modular repository to train ML algorithms - and especially Generative Adversarial Networks (GANs) - on electrocardiographic data, even though the setup is generally suitable for arbitrary time series. While ECGAN also supports other tasks, such as classification using simple RNNs/CNNs, the focus lies on the generation of data using GANs. The GANs are trained on only the ''normal'' (for ECGs: healthy) class to learn the underlying representation of this class. The resulting GAN can be used to generate new healthy data or detect anomalies by comparing the synthetic data with the test data. ECGAN follows modular setup allows easy modifications to the different parts of this project, allowing researchers to simply download and preprocess various ECG datasets, train a variety models to generate realistic time series and apply a variety of anomaly detection mechanisms. ECGAN is built to be reproducible and easily expandable to allow fast prototyping and foster comparability and open source implementations by reducing the time required to implement models or preprocessing scenarios.
Parts of the implementation are still in an experimental state.
Make sure to have installed python3.8 and git.
- Clone the repository (using
git clone) - Setup the virtual environment (e.g.
make envwhich executesvirtualenv --python=/usr/bin/python3.8 .venvand activate the environmentsource .venv/bin/activate) - Run
make setup_devin root if you want to develop. Runmake setupif you only want to execute experiments. - During Step 2, ecgan and all requirements will be installed.
If you want to use ecgan on Windows, you cannot use make without additional tools. You can either use chocolatey
or similar tools to install make and use it. Otherwise, you can manually recreate the steps:
Install virtualenv (pip install virtualenv), create the environment virtualenv --python=python3.8 .venv and
activate the env (.venv\Scripts\activate) and install the packages (pip install -U -r requirements.txt)
The package comes with a convenient CLI and configuration over YAML-Files.
By calling ecgan-init -d DATASET -m MODULE -o FILE_NAME.yml ENTITY PROJECT EXPERIMENT_NAME a configuration file
is generated. This can be manually updated and contains most relevant hyperparameters.
After possibly applying changes to the config file (lets say FILE_NAME) you can start the preprocessing of
the chosen dataset using ecgan-preprocess FILE_NAME.yml before training the model by invoking ecgan-train FILE_NAME.yml.
To perform anomaly detection, the -a model_reference flag is used (see docs for more
information).
We support several datasets will be used for evaluation: MIT-BIH (small amount of patients, long recordings, often used), the PTB ECG dataset as well as the Shaoxing dataset (many patients, smaller recordings, published recently).
Some of the datasets require setting up kaggle credentials or downloading it manually (see docs)
ECGAN is an initiative of eMundo GmbH and the LMU Munich research group data mining in medicine supported by the Bavarian Research Foundation. The purpose of this initiative is to investigate anomaly detection in time series data, focusing on electrocardiogram data. Do not hesitate to reach out if you have any questions.