A hybrid deep learning algorithm for ECG-based heart disease classification

Abstract Automated electrocardiogram (ECG) classification for heart disease detection encounters significant challenges, such as patient variability, inter-patient differences in heart signals, and data leakage when information from the same patient is present in both training and test sets. To mitigate these challenges, a Hybrid 1D CNN-BiLSTM (one- dimensional convolutional and bidirectional long short-term memory) deep learning framework is introduced, employing rigorous patient-wise validation on the MIT-BIH Arrhythmia Database. Patient-wise data splitting (34 records for training, 14 records for testing) ensures that each patient’s data is confined to a single set, effectively preventing data leakage. Under these conditions, the proposed Hybrid with GAN model achieves 91.69% accuracy (95% confidence interval: 91.41−91.99%), with a macro F1-score of 0.547 ± 0.374. Conditional GAN augmentation significantly improves minority class performance: Supraventricular recall increases from 9.2% to 20.9% (+127.2%), and Ventricular recall improves from 77.7% to 80.9% (+4.1%). Statistical significance testing using McNemar’s test demonstrates the superiority of Hybrid with GAN over the non-augmented hybrid model ( $$\chi ^2 = 12.41$$ , p = 0.0004). The observed 7.31% performance gap between beat-wise (99.0%) and patient-wise (91.69%) evaluation highlights the critical importance of patient-wise validation for clinically realistic performance estimates. Overall, this framework provides a rigorous patient-wise evaluation protocol for automated ECG classification, with patient-wise results for a GAN-augmented hybrid model on the full 5-class MIT-BIH dataset.