An optimized convolutional neural network for arrhythmia classification
<p>Electrocardiogram (ECG) is a practical medical test to diagnose arrhythmia. As a</p><p>crucial computational application in clinical practice, ECG automatic classification can</p><p>effectively detect the possible occurrence of...
Saved in:
| Main Author: | |
|---|---|
| Format: | thesis |
| Language: | zsm |
| Published: |
2022
|
| Subjects: | |
| Online Access: | https://ir.upsi.edu.my/detailsg.php?det=9598 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | <p>Electrocardiogram (ECG) is a practical medical test to diagnose arrhythmia. As a</p><p>crucial computational application in clinical practice, ECG automatic classification can</p><p>effectively detect the possible occurrence of cardiovascular disease. At present, the</p><p>main problems in the automated classification of ECG are due to (1) the complexity of</p><p>algorithms to capture heartbeats, (2) the complex changes of irregular heartbeats in</p><p>rhythm or morphology leading to difficulties in the ECG feature recognition, and (3)</p><p>the needs of large training samples and training time for a machine learning to achieve</p><p>the ideal recognition accuracy. Given the problems in ECG automatic classification,</p><p>this study proposes an effective automated classification approach for arrhythmia based</p><p>on a representative convolution neural network that can decode ECG source files and</p><p>identify heartbeats accurately based on the detection of QRS waveform from the ECG</p><p>records. A one-dimensional convolutional neural network (1D-CNN) is proposed to</p><p>accurately classify different types of arrhythmias by automatically extracting the</p><p>morphological features of ECG. The initial connection weights of 1D-CNN are</p><p>optimized based on differential evolution to improve its ECG classification. The</p><p>optimized 1D-CNN is evaluated against two arrhythmia databases, namely the MITBIH</p><p>and SCDH arrhythmia databases. Besides, a comparison is made between the</p><p>optimized and unoptimized 1D-CNN. The results show that the proposed model has</p><p>higher accuracy in heartbeat classification. Compared to the unoptimized 1D-CNN, the</p><p>accuracy improves by 0.6% and 3.1%, respectively. Besides, the optimized 1D-CNN</p><p>requires less training time, 9.22 seconds less with MIT-BIH and 10.35 seconds less with</p><p>SCDH based on ReLU active function and 10 epochs, as compared to the unoptimized</p><p>1D-CNN based on the same parameter settings. The training time of the optimized 1DCNN</p><p>decreased by 67.2% and 64.2% with MIT-BIH and SCDH, respectively.</p> |
|---|