ECG Heartbeat Classification Using CNN Autoencoder Feature Extraction and Attention-Augmented BiLSTM Classifier

The electrocardiogram (ECG) is a primary non-invasive technique for recording the heart’s electrical activity, but the complexity and subjectivity of manual interpretation have driven the growth of computer-aided diagnostic systems to improve accuracy and reliability in detecting cardiac abnormalities. This research proposes a deep learning (DL)-based framework that integrates a convolutional autoencoder (CAE) for dimensionality reduction and unsupervised feature extraction from raw ECG signals. The pipeline then incorporates a Bidirectional Long Short-Term Memory (BiLSTM) network enhanced with a multi-head self-attention mechanism for robust and automated heartbeat classification. To validate this architecture, experiments were conducted using the MIT-BIH Arrhythmia dataset focusing on five arrhythmia classes. Additionally, we benchmark our system’s performances against state-of-the-art methods. Our results demonstrate the efficiency of this method with an overall accuracy of 98.57%.