Project Raccoon

This is part of my master's thesis. I evaluate how neural networks perform on the task of finding QRS complexes in single-channel ECG signals of varying signal quality. The name (raccoon) has no further meaning, I just like raccoons and needed a name. Think of a zealous raccoon searching for QRS complexes!

What is this?

I compare how different detectors (algorithms searching for QRS complexes) perform. This is done by using the detectors on ECG signals and comparing the QRS complex positions found with the positions of actual QRS complexes in the signal. Some of these detectors employ neural networks and need a training step, some others employ means of signal processing and do not need training. Whilst the first are subject of this work the latter one establish a baseline to compare the performance of neural networks with.

How to use this?

For using this, three steps are to be performed:

Install requirements
Download databases
Run evaluation

Install Requirements

You need Python3 and PIP. Everything else is easily done with:

pip install -r requirements.txt

Download Databases

We provide a Python script for doing this. Call scripts/download_databases.py and specify which databases you want to get.

We support:

MIT-BIH Arrhythmia Database (mitdb)
MIT-BIH Normal Sinus Rhythm Database (nsrdb)
MIT-BIH Noise Stress Test Database (nstdb)

For example, for downloading the mitdb to a directory called data:

scripts/download_databases.py -d data -k mitdb

For more noise contaminated data use scripts/noisy_mitdb.py. This script uses nst from the WFDB software package to create noisy records from all mitdb records using the electrode motion em noise template from nstdb.

scripts/noisy_mitdb.py # writes records to current working directory

Run Evaluation

Call run.py specifying a configuration file. Configurations files tell the evaluator what to do. You can find example configuration files in the configurations directory. Feel free to change them according to needs and research questions. Parameters specified in the configuration file are the same the constructors of Evaluator and Detector classes take, so you can use them as reference.

Example call:

./run.py configurations/example01.json

Run in a Docker Container

The Dockerfile is made for building Docker images to run on GPU machines. Hence, nvidia-docker is required to build images. In the root directory of this repository run:

nvidia-docker build -t raccoon:latest .

To start a container running this image:

nvidia-docker run -it raccoon bash

From there you can call run.py as described above.

Which detectors do we compare?

We use three NN-based detectors and three signal processing based detectors.

The NN-based detectors use neural network architectures proposed by the following papers:

C. García-Berdonés, J. Narváez, U. Fernández, and F. Sandoval, “A new QRS detector based on neural network,” in Biological and Artificial Computation: From Neuroscience to Technology, 1997, pp. 1260–1269.
M. Šarlija, F. Jurišić, and S. Popović, “A convolutional neural network based approach to QRS detection,” in Proceedings of the 10th International Symposium on Image and Signal Processing and Analysis, 2017, pp. 121–125.
Y. Xiang, Z. Lin, and J. Meng, “Automatic QRS complex detection using two-level convolutional neural network,” Biomed Eng Online, vol. 17, Jan. 2018.

The signal processing based detectors are GQRS and XQRS by MIT's wdfb Python module and a handwritten version of the Pan-Tompkins algorithms.

Cf.: J. Pan and W. J. Tompkins, “A Real-Time QRS Detection Algorithm,” IEEE Transactions on Biomedical Engineering, vol. BME-32, no. 3, pp. 230–236, Mar. 1985.

How do we compare?

For each ECG signal, we do have a list of positions telling us where the QRS complexes actually are. The detectors generate lists of positions where QRS complexes where detected. We compare these lists and produce the classification metrics true positives, true negatives, false positives, and false negatives. From this we can compute further classification metrics such as sensitivity, F1 score, and positive predictive value.

This is done by applying the following rules:

A detected QRS complex counts as true positive if there is an actual QRS complex within a given range.
A detected QRS complex counts as false positive is there is no actual QRS complex within a given range.
An actual QRS complex having no detected QRS complex within a given range counts as false negative.
There are no true negatives.

gabriel19913/raccoon