The electrocardiogram (ECG) is an efficient tool to assess heart health and diagnose heart arrhythmias (irregular heartbeats) by capturing the heart’s electrical activity. Atrial Fibrillation (AF) is one the most common serious arrhythmias and is therefore of particular interest to diagnose.

We focus on training a classifier to automatically detect if an arrhythmia is present in an ECG and distinguish whether or not the arrhythmia is AF. To accomplish this, we first preform 6-level wavelet decomposition with a Symlet5 wavelet on each ECG signal to create two new corresponding time series. One of the time series will isolate the QRS complex component of the ECG and the other will isolate the T-wave and P-wave components of the ECG (these components play a key role in diagnosing arrhythmias).

For each ECG in our training set, we combine the original ECG signal with the two component isolating time series mentioned above. We will then train a multi-channel, deep convolutional neural network (MC-CCN) directly from this multivariate time series data and use it to classify whether or not AF or another arrhythmia is present. Unlike many traditional statistical models this model will take advantage of automatic feature building, reducing the need for extensive domain expertise normally require for feature construction.

We will train our model on a labelled data set (collected from the portable AliveCor ECG device) containing the following 4 classes: normal, AF, other arrhythmia, or noisy. The data set contains 8528 single lead ECG recordings lasting from 9 to 60 seconds, each of which were sampled at 300 Hz. Performance will be evaluated with a 10-fold cross validated F1 score. We will perform Wavelet decomposition in Python with the Pywavelet package and the model building will be done in Python and R.


Author: Alexander Barrett

Coauthor(s): Alexander Barrett

Status: Work In Progress