Comparing Deep Neural Network for Multi-Label ECG Diagnosis From Scanned ECG

Electrocardiograms (ECGs) play a vital role in diagnosing cardiovascular diseases (CVDs), which remain one of the leading causes of mortality worldwide. The accurate interpretation of ECG signals is crucial for early detection and timely medical intervention. Recent advancements in deep learning have significantly improved ECG-based diagnosis, with models achieving cardiologist-level performance [1, 2, 3, 4, 5, 6, 7]. However, most of these approaches rely on high-quality digital ECG data, limiting their real-world applicability in clinical environments where scanned paper ECGs are still prevalent. Paper-based ECGs remain widely used due to historical adoption, cost-effectiveness, and compatibility with legacy healthcare systems. However, relying on scanned ECGs introduces new challenges, as they contain image-based artifacts such as noise, distortions, and variations in paper quality, which can affect automated diagnostic accuracy. Traditional binary classification (normal vs. abnormal) methods may not fully capture the complexity of cardiac conditions present in real-world ECGs. Therefore, multi-label classification, where multiple cardiac abnormalities are identified simultaneously, presents a more clinically relevant and challenging problem. Recent efforts in ECG analysis have explored deep neural networks, including AlexNet, VGG, ResNet, and Vision Transformers, for automated classification of ECGs.