Deep Learning Approach for Detection of Atrial Fibrillation and Atrial Flutter Based on ECG Images

Abstract

This study explores the application of image-based deep learning techniques to distinguish between Atrial Fibrillation (AFib) and Atrial Flutter (AFlut) using images of standard 12-lead ECG exams from a private database. By implementing a MobileNet Convolutional Neural Network architecture, we achieve a high classification performance, with an accuracy of 95.6%, AUROC of 97.6%, F1-score of 83.2%, specificity of 99.6%, and sensitivity of 72.7%. We also applied explainable methods, such as Grad-CAM and LIME, to try to interpret the model’s decision-making process and identify significant regions within the ECG images that contribute to the classification. Our results demonstrate the potential of image-based deep learning approaches for accurate and reliable discrimination between AFib and AFlut, paving the way for enhanced diagnostic capabilities in clinical settings.