README.md

Coronary ARtery disease PrEdictor (CARPE)

This is the official repository for the paper
Enhancing the diagnosis of functionally relevant coronary artery disease with machine learning.

Prerequisites

Before cloning the repository, make sure you have git lfs installed. This is necessary to be able to check out the model checkpoints exceeding 100MB due to github's file size limit.

Once git lfs is installed, clone the repository and install all dependencies with pip install -r requirements.txt. The code is tested with Python 3.8.

Generating Predictions

Take a look at our sample notebook to learn how to use $CARPE_{\text{Clin.}}$, our random forest trained on a small set of static clinical data, and our neural network approach $CARPE_{\text{ECG}}$ which takes both ECG signals and static date as inputs.

Data Preprocessing

To preprocess your custom ECG signals, you will have to write your own data loader depending on your file format. We recommend inheriting from THEWParser . The main function you have to implement is _get_raw which loads the raw ECG signal according to your data format. Loading should result in a numpy array of dimensions [T, num_leads], where $T$ is the length of the signal. The sampling rate of you signal should be either 500Hz or 1000Hz (take a look at the paper for more details). Once you can load your data into your custom parser the following code snippet applies all preprocessing steps that we used in the manuscript.

import numpy as np
parser = THEWParser(filepath) # Replace with your parser

# Preprocess
band = [0.05, 150.0]
parser.apply_butter(parser.data, [band[0]/(parser.freq/2), band[1]/(parser.freq/2)])
parser.apply_median(parser.data)
parser.apply_smoothing(parser.data)
parser.apply_winsorizing(parser.data, 0.05, 100 - 0.05)

downsampled = signal.decimate(parser.data, 2, axis=0)

np.savez(OUTPUT_PATH, data=downsampled.T)

2-6-2 Sequence Extraction

You will likely not have access to the exact times when the stress phase started/ended. Instead, you can use the the time point of the maximum heart rate as the last point from which stress windows are extracted. To extract the first 2-6-2 sequence, take the first 2 seconds of the ECG signal, the last 6 seconds that preceed the time point with maximal HR, and the last 2 seconds of the ECG signal. For the second 2-6-2 sequence, stride the first window forward by 2 seconds, the second window back by 6 seconds, and the third window back by 2 seconds. Continue until you extracted all 2-6-2 sequences. Take a look at this function for an implementation.