Classification of ECG Signals for Detecting Coronary Heart Diseases Using Deep Transfer Learning Techniques

Abstract

Deep learning (DL) techniques were applied to ElectroCardioGrams (ECGs) for identifying abnormalities in heart diseases. The application of a convolutional neural network is discussed in this paper to perform the detection of coronary heart diseases by extracting QRS complexes and the error signal features. AlexNet, VGG19, ResNet50, GoogleNet, and NasNetLarge pretrained transfer learning models are adopted using the adaptive moment estimate (Adam) and stochastic gradient descent with momentum (SGDM) optimizers for training, testing, and validation. Here, 3 classes of heart diseases have been considered; namely arrhythmia, cardiomyopathy, and ischemia. The suggested method is aimed at automatically detecting these diseases based on ECG signals collected from MIT-BIH databases. The obtained results show that the Adam optimizer outperforms the SGDM optimizer for the five DL architectures. For AlexNet adopting Adam, the accuracy of detecting Arrhythmias, Ischaemia, and Cardiomyopathy is 98.2%, 95.9%, and 93.5% respectively. ResNet-50 and NasNetLarge with the same optimizer, have 98.0%, 96. 9%, 92.3 %, 97.9%, 95.7%, and 94.0 % accuracy in detecting Arrhythmias, Ischaemia, and Cardiomyopathy respectively. In addition, the subtraction of the QRS complexes from the clean ECG signal computationally outperforms the cutting-edge method based on using the continuous wavelet transform method. The reason is that the wavelet method is computationally expensive compared to the proposed QRS-complex subtraction method that results in integer error samples. So, the average execution time is significantly less.