Abstract:
Electrocardiograms (ECGs) can be used as a reliable method of measurement
for monitoring the functions of the heart and the circulatory system. There has been an
increased emphasis on the accurate identification and classification of heartbeats in
recent times. The process of measuring specific characteristics of signal is feature
extraction. This process reduces the original data as necessary and aims to distinguish
between different input patterns. Feature vectors are extracted using the wavelet
transform. In this study, Daubechies (db4, db6, db16) and Symlet wavelets are
employed within the Discrete Wavelet Transform while the Morlet wavelet is used in
the Continuous Wavelet Transform, to facilitate a comparative analysis of wavelet
coefficients and their accuracy. During the preprocessing stage, this model incorporates
band-pass filtering and normalization techniques. Furthermore, it utilizes the wavelet
transform for feature extraction and integrates CNN attention mechanisms within the
classifier model. The feature extraction method employing the db16 wavelet
demonstrates higher accuracy achieving a classification accuracy of up to 97% when
comparing with the other types of wavelets int the MIT-BIH dataset. “Balanced class
weights” have been incorporated to address the issue of class imbalance within the
dataset.
Conducting experiments on the MIT-BIH ECG dataset and analyzing model
performance across various approaches, including single-lead ECG, multi-lead ECG,
lead fusion, and feature fusion, using the wavelet attention technique. The proposed
method offers high classification capability and can effectively identify cardiac
arrhythmias according to the results. The lead fusion method achieved a classification
accuracy of 0.98. In the evaluation of single-lead, multi-lead, and feature fusion
experimental approaches, classification accuracies of 0.97, 0.98, and 0.97 were
obtained respectively. The model that combined the class-weighting method with latefusion feature extraction achieved an accuracy of 0.95. Meanwhile, other methods
involving class weighting yielded accuracies of 0.91, 0.92, and 0.92 respectively, and
a finding that demonstrates how model performance can vary depending on the specific
approach used. An ECG classification framework focuses on cross-dataset learning by
combining various publicly available ECG datasets specifically, MIT-BIH and PTBXL, to create a unified training set. The proposed approach addresses variations in signal characteristics, sampling
frequencies, noise levels, and patient demographics by employing standardized
preprocessing, normalization, data harmonization, resampling, and wavelet transform
techniques. A deep learning model has been trained to learn general representations of
ECG patterns. Utilizing a Wavelet Convolutional Neural Network (CNN) with a selfattention mechanism, the model classifies ECG signals into five categories: NORM
(normal), CD (conduction disturbance), V (ventricular), S (supraventricular), and
"Other". Experiments conducted on the MIT-BIH arrhythmia dataset, the PTB-XL
dataset, and a combined dataset have been analyzed. In the evaluation of multiclass
classification tasks, the models achieved accuracy rates of 98.19% for the MIT-BIH
dataset, 84.18% for the PTB-XL dataset, and 94.36% for the combined dataset.