ECG Classification for Detecting ECG Arrhythmia Empowered with Deep Learning Approaches

Abstract

According to the World Health Organization (WHO) report, heart disease is spreading throughout the world very rapidly and the situation is becoming alarming in people aged 40 or above (Xu, 2020). Different methods and procedures are adopted to detect and diagnose heart abnormalities. Data scientists are working on finding the different methods with the required accuracy (Strodthoff et al., 2021). Electrocardiogram (ECG) is the procedure to find the heart condition in the waveform. For ages, the machine learning techniques, which are feature based, played a vital role in the medical sciences and centralized the data in cloud computing and having access throughout the world. Furthermore, deep learning or transfer learning widens the vision and introduces different transfer learning methods to ensure accuracy and time management to detect the ECG in a better way in comparison to the previous and machine learning methods. Hence, it is said that transfer learning has turned world research into more appropriate and innovative research. Here, the proposed comparison and accuracy analysis of different transfer learning methods by using ECG classification for detecting ECG Arrhythmia (CAA-TL). The CAA-TL model has the multiclassification of the ECG dataset, which has been taken from Kaggle. Some of the healthy and unhealthy datasets have been taken in real-time, augmented, and fused with the Kaggle dataset, i.e., Massachusetts Institute of Technology-Beth Israel Hospital (MIT-BIH dataset). The CAA-TL worked on the accuracy of heart problem detection by using different methods like ResNet50, AlexNet, and SqueezeNet. All three deep learning methods showed remarkable accuracy, which is improved from the previous research. The comparison of different deep learning approaches with respect to layers widens the research and gives the more clarity and accuracy and at the same time finds it time-consuming while working with multiclassification with massive dataset of ECG. The implementation of the proposed method showed an accuracy of 98.8%, 90.08%, and 91% for AlexNet, SqueezeNet, and ResNet50, respectively.