Heart Arrhythmias Detection and Classification using ECG Signals and Deep Convolutional Neural Networks
No Thumbnail Available
Date
2021-12
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
According to the World Health Organization, almost 17 million people die each year as a
result of cardiovascular illness. The irregularity and abnormalities of heartbeat rhythm which
is known as arrhythmia is one of the conditions that can affect the cardiovascular system.
Electrocardiogram (ECG) is a reliable tool that can be used for monitoring the cardiovascular
health. Recently, classifying the ECG signals based on Artificial Intelligence (AI) is
increasingly being studied. Convolutional Neural Networks (CNN) in particular have been
effectively applied for the classification of ECG signals. Although high prediction accuracies
have been reported, majority of previous studies have only been developed to classify limited
number of arrhythmias. The methods were developed to evaluate all major types of
arrhythmias using 1-D CNN to classify time domain representation of ECG waveforms.
However, using 1-D CNNs has limited flexibility due to the use of 1-D kernels. There are
methods reported to transform the time series signals into 2-D images using STFT and use 2-
D CNN. However, STFT is difficult to apply to non stationary signals; there is no way to
resolve the complete frequency content of such signals with a single localizing window size.
To overcome this obstacle of Fourier decomposition, the Continuous Wavelet Transform
(CWT) could be used to breakdown a signal into wavelets with a high degree of temporal
localization. The S-transform could be another option since it takes the advantage of STFT
and wavelet. This thesis study uses CNN classifiers for detecting and classifying heart
arrhythmias based on analysis of ECG signals in time-frequency domain. The used data were
extracted from a subset of MIT-BIH arrhythmia data set, that contain 1000 ECG signals of 17
classes in total, collected from 45 patients. 12 classes were chosen from the subset which
include Normal Sinus Rhythm (NSR), Atrial Premature Beat (APB), Atrial Flutter (AFL),
Atrial Fibrillation (AFIB), Supraventricular Tachyarrhythmia (SVTA), Premature Ventricular
Contraction (PVC), Ventricular Tachycardia (VT), Idioventricular Rhythm (IVR),
Ventricular Flutter (VFL), Left Bundle Branch Block Beat (LBBB), Right Bundle Branch
Block Beat (RBBB), and Pacemaker Rhythm (PR). The one dimensional ECG signals were
transformed into joint time frequency spectrograms using Stockwell transform and into
scalograms using Continuous Wavelet Transform (CWT). By using different pretrained
networks for classifying spectrograms and scalograms namely GoogleNet, SqueezeNet, and
ResNet-50, different results were achieved. GoogleNet pretrained network showed the best
v
performance when using CWT generated scalograms with 93.85% accuracy, 96.42%
precision, 84.14% sensitivity, 99.36% average specificity and 89.86 F1-score. Based on the
results, transfer learning especially GoogleNet proved to be efficient in classifying the twodimensional
scalograms of cardiac arrhythmias, while reducing the burden of training
network from scratch makes it easily applicable. Compared with recent techniques, results
obtained using the proposed technique show the great promises of the 2-D CNN model in
accurate classification of arrhythmias using CWT and S-transform and the proposed method
resulted in higher accuracy and F1-score.
Description
Keywords
Cardiac Arrhythmias, Electrocardiogram, Deep Learning, Convolutional Neural Network, Stockwell Transform, Continuous Wavelet Transform