Electrocardiogram Signal Classification for Heart Disease Diagnosis with Deep Learning A Sri Lankan Study

Abstract

Heart disease remains a significant health concern worldwide, including in Sri Lanka, where timely and accurate diagnosis is crucial for improving patient outcomes. The 12-lead electrocardiogram (ECG) is a widely used, cost-effective tool for detecting cardiac abnormalities. This study presents a deep learning-based approach for classifying 12-lead ECG images into normal and not normal categories using a region-specific dataset collected from Jaffna Teaching Hospital in Sri Lanka. ECG records were digitized from printed paper using a mobile application that converts physical documents into high-quality digital images. The dataset was organized into three formats: scanned images, digitized signals in Comma-Separated Values format, and clean ECG traces without grid lines. The preprocessing pipeline focused on creating standard-sized 3×4 combined lead images through segmentation, binarization, and noise removal. These images served as inputs for convolutional neural network architectures such as ResNet-18, SqueezeNet1.1, MobileNetV2, EfficientNet-B1, and a custom BasicDenseNet. The Convolutional Neural Network models were trained end-to-end using 12-lead ECG image inputs. Each lead was treated as a separate channel, enabling the model to learn spatial and morphological dependencies across multiple leads. The network automatically extracted discriminative features such as waveform shapes, lead-specific variations, and inter lead correlations that influenced the classification outcome. Models were trained and evaluated to optimize classification accuracy and generalization. Model performance was assessed using accuracy, precision, recall, F1-score, and confusion matrices to ensure balanced evaluation. ResNet-18 achieved the highest validation accuracy of 93%, followed by SqueezeNet1.1 at 90%. To enhance interpretability and clinical trust, Gradient weighted Class Activation Mapping visualizations were used to identify discriminant ECG regions influencing predictions. These visualizations, reviewed by cardiographers, confirmed that the model focused on clinically relevant leads and waveform patterns. This study demonstrates the potential of combining deep learning with explainable AI to build reliable ECG classification systems. It highlights the effectiveness of deep learning for ECG-based abnormality detection, providing a scalable pipeline to aid early diagnosis of heart abnormalities in the Sri Lankan population.