Cardiovascular Classification Using Efficient Net on Electrocardiogram Images

Abstract

Cardiovascular disease ranks among the top causes of mortality, frequently caused by sudden obstructions within blood vessels. Timely identification and intervention are essential for minimizing the impact of the disease. This research employs image augmentation techniques to correct class imbalance in an ECG image dataset divided into five categories: Normal, Abnormal Heartbeat, Myocardial Infarction, Previous History of Myocardial Infarction, and COVID-19. The balanced dataset includes 6,322 images. To improve classification accuracy for cardiovascular diseases, three pre-trained models visual Geometric Group, Residual, Dense, and Efficient Network with Version 2, were trained on the balanced ECG dataset. Critical hyper parameters were fine-tuned, yielding optimal performance with a learning rate set at 0.00001, a dropout rate of 0.3, and utilizing the Adam optimizer. EfficientNet-V2 outperformed the other models, reaching a level of accuracies of 96.22%, precision 96.34%, recall 96.31%, 95.89%, 94.75%, and an F1-Score of 96.33%, thus exceeding the performance of Densenet 161, Densenet 201, ResNet50 and VGG16.