This repository contains the code to train a convolutional neural network to classify the gender of a patient from their ECG signal

Clone the repository and install the project and its dependencies. Numpy needs to be installed first because it is required by pyedflib.

git clone https://github.com/FeliMe/ECG-gender-classification.git
pip install numpy
pip install -r requirements.txt

You need to apply for the following datasets (SHHS, MESA), download them and convert them to .npy files using

data/convert_data.py

• For the SHHS dataset, only the follow-up clinic visit (SHHS2) has the right sampling rate. Navigate to shhs/files/polysomnography/edfs/shhs2 to get those files.
• The MESA the files can be found at mesa/files/polysomnography/edfs
• In addition, you need to download the following files which include the meta-data: SHHS-Meta MESA-Meta

Before running the convert_data.py, make sure to adapt the paths in the config section accordingly.

Adapt the config section in train.py and run the script

python3 train.py

We report a 74% accuracy on the test set.

Evaluation of the trained model is done in the jupyter notebook eval.ipynb.

Saliency maps show which part of the input contributes the most to the decision made by the network. The two maps show input where high saliency scores are marked by red and blue dots.

The limitations of this project are mainly due to the data available. Further work could include using data from more patients and with more ECG-channels. It would also be interesting to predict medically more useful features than the gender of a patient.