A Deep Learning Framework for Automatic Cardiovascular Classification from Electrocardiogram images (preprint)

ABSTRACT

Cardiovascular disease is the primary reason of death within a short period because of a sudden blockage in blood vessels. Therefore, early diagnosis and treatments are necessary to reduce the disease severity. In this study, Image augmentation concepts are used for balancing the ECG images dataset classes. The balanced dataset contains 6322 images from five different classes are Normal, Abnormal Heartbeat, Myocardial Infraction (MI), Previous History of MI, and Covid-19 ECG images. The pre-trained models VGG-16, ResNet-50, and DenseNet-161 are used to train on ECG balanced dataset. The use of hyper-parameters will increase the accuracy of classifying cardiovascular disorders which includes learning rate, optimizer, and dropout. The mentioned pre-trained models reported the highest accuracy at learning rate = 0.00001 and dropout=0.3 with Adam optimizer. DenseNet model achieved 93.33% accuracy, 94.23% precision, 93.33% recall, and 93.30% F1-Score, which is higher in performance compared to ResNet-50 and VGG-16 pre-trained models.