A lightweight deep learning approach for patient-specific electrocardiogram beat classification using local and long-term dependencies

An electrocardiogram (ECG) is a graphical tool used to assess patients’ cardiac activity. Long-term ECG recordings, typically spanning 24 to 48 h, are crucial for detecting cardiac disorders. This paper introduces a novel, lightweight deep-learning architecture for classifying ECG beats as per the AAMI (Association for the Advancement of Medical Instrumentation) standard. The model integrates the advantages of Convolutional Neural Networks (CNN) and Bi-directional Long Short-Term Memory (Bi-LSTM) mechanisms in a single network to effectively capture local, temporal, and sequential patterns in ECG signals. Unlike conventional training, which often relies on fixed learning rates or predefined epochs, the proposed method dynamically adjusts learning parameters based on validation performance. Two Bi-LSTM layers effectively capture rich temporal dependencies, without requiring additional depth. The proposed method concatenates extracted CNN and BiLSTM features before the compact, dense layer, which will reduce the number of parameters significantly. This lightweight model ensures fast inference and low computational costs. Experimental results show that the proposed method achieves an accuracy of 99.21%, sensitivity of 98.66%, precision of 99.19%, and an F-score of 0.987. Additionally, the model demonstrates strong generalization capabilities, achieving high accuracies of 96.17% over different databases. The model‘s robustness and reliability in classifying ECG beats make it a practical and efficient tool for real-time monitoring applications.