Authors - Ei Marlar Win, Amy Tun, Khant Kyawt Kyawt Theint Abstract - Electrocardiogram (ECG) signal analysis plays an important role in the early detection and diagnosis of cardiovascular diseases. Manual interpretation of ECG recordings is time-consuming and highly dependent on clinical expertise, creating a need for automated and accurate classification systems. This study presents an automated ECG classification model using signal preprocessing, heartbeat segmentation, wavelet, feature extraction, and deep learning. ECG signals are preprocessed to remove noise using filtering and normalization methods. Features are extracted heartbeat segments-based windows around each R peak and classified into five different arrhythmias N (Normal), V (Ventricular), S (Supraventricular), F (Fusion) and Q (Unknown/noisy /unclassified) using wavelet Convolutional Neural Network (CNN) Self Attention model. Experiments on MIT-BIH ECG dataset and analyze the model performance evaluation across a single-lead ECG, multi lead ECG, lead fusion and feature fusion techniques by wavelet attention. The results indicate that the proposed approach yields high classification performance and effectively distinguishes heartbeats abnormalities. Class weighting techniques were applied to address the issue of imbalanced class labels in the ECG dataset. The lead fusion approach achieved classification accuracies of 0.98. Single lead, multi lead and feature fusion experimental approaches were evaluated, resulting in classification accuracies of 0.97, 0.98, and 0.97, respectively. The class-weighting method combined with lead fusion feature extraction obtained an accuracy of 0.95. Furthermore, class weight additional techniques achieved accuracies of 0.91, 0.92 and 0.92, demonstrating variations in model performance across different methodologies. This automated system can support clinicians to assist in the early diagnosis of heart abnormalities and improve healthcare efficiency.
Open Zoom