Clone the repository

git clone [email protected]:manideep2510/ECG-acquisition-classification.git

Python 3.5 or higher is required to run the code. To test the code the with the pretrained models, Python 3.5 is the only one it supports

Install the requirements (this may take a few minutes).

For CPU only support run

cd ecg
./setup.sh

To install with GPU support run

env TF=gpu ./setup.sh

In the repo root direcotry (ecg) make a new directory called saved.

mkdir saved

To train a model use the following command, replacing path_to_config.json with an actual config:

python ecg/train.py path_to_config.json

Note that after each epoch the model is saved in ecg/saved/<experiment_id>/<timestamp>/<model_id>.hdf5.

For an actual example of how to run this code on a real dataset, you can follow the instructions in the cinc17 README. This will walk through downloading the Physionet 2017 challenge dataset and training and evaluating a model.

After training the model for a few epochs, you can make predictions with.

python ecg/predict.py <dataset>.json <model>.hdf5

replacing <dataset> with an actual path to the dataset and <model> with the path to the model.

This work is published in the following paper in Nature Medicine

Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network

If you find this codebase useful for your research please cite:

@article{hannun2019cardiologist,
  title={Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network},
  author={Hannun, Awni Y and Rajpurkar, Pranav and Haghpanahi, Masoumeh and Tison, Geoffrey H and Bourn, Codie and Turakhia, Mintu P and Ng, Andrew Y},
  journal={Nature Medicine},
  volume={25},
  number={1},
  pages={65},
  year={2019},
  publisher={Nature Publishing Group}
}