Heart Disease Prediction Using EfficientNet-B0: A Deep Learning-Based ECG Image Analysis Approach

Abstract

Autonomous Electrocardiogram (ECG) image classification contributes to the effective detection of cardiac abnormalities and decrease the reliance on manual interpretations. This research study integrates transfer learning and EfficientNet-B0 architecture for the classification of four-class ECG images. In the pre-trained model, the open-source ECG images are pre-processed via image size adjustment, tensor conversion, and channel wise normalization. EfficientNet-B0 architecture combined with the weights obtained via ImageNet training was fine-tuned by enabling a task specific classifier with four output classes and training with cross-entropy loss function and Adam optimizer. The model performance analysis is done in terms of accuracy, precision, recall, F1-score, and confusion matrix analysis. The resultant ECG image classification corresponds to the accuracy of 99.61%. The class-level performance demonstrates high model prediction results. Moreover, the proposed framework performs efficient classification and visualization of individual ECG images.