A deep hybrid CNN–BiGRU model for inter-patient ECG beat classification

Abstract

This study aims to develop and validate a robust machine learning-based framework for automated ECG beat classification, with the purpose of supporting rapid arrhythmia detection and improving the efficiency of clinical decision support systems. The work focuses on assessing the effectiveness of hybrid deep learning architectures that can jointly capture local morphological patterns and sequential dependencies in ECG signals. The proposed experimental framework was designed using publicly available MIT-BIH Arrhythmia and AAMICS INCART databases. ECG recordings were pre-processed and augmented through Gaussian noise addition, amplitude reduction, and temporal shifting to improve model generalization under signal variability. Several deep learning models were trained and evaluated, including convolutional and hybrid recurrent architectures, to identify the most effective approach for ECG record analysis. The findings demonstrate that the CNN–BiGRU architecture achieved the best overall performance, with an accuracy of 93.89%, precision of 93.36%, recall of 93.91%, and F1-score of 93.39%. These results indicate that combining convolutional neural networks with bidirectional gated recurrent units improves the representation of both morphological and temporal ECG characteristics. The study provides proof of concept for using hybrid deep learning architectures in automated ECG analysis. The practical implication of this work is that the proposed CNN–BiGRU framework may support faster, more reliable ECG interpretation and has potential applicability in real-time arrhythmia screening, wearable monitoring systems, and clinical decision support environments.