Abstract
Accurate detection of cardiac arrhythmias from electrocardiogram (ECG) signals is essential for the early diagnosis of cardiovascular diseases but remains challenging due to the complex, non-linear nature of ECG waveforms. This study proposes HCTG-Net, a Hybrid CNN-Transformer Network with Gated Fusion, designed to jointly capture local morphological features and long-range temporal dependencies in ECG data. The model employs a dual-branch architecture, where a residual CNN extracts localized waveform patterns and a Transformer branch models global temporal context. A learnable gated fusion mechanism adaptively balances and integrates features from both branches at the per-dimension level. Experiments conducted on the MIT-BIH Arrhythmia Database demonstrate that HCTG-Net achieves superior performance compared with existing methods, reaching an overall accuracy of 0.9946 and F1-score of 0.9711. Visualization results show well-clustered feature distributions, confirming robust feature learning, while ablation studies verify the complementary roles of the CNN, Transformer, and fusion modules. Overall, HCTG-Net offers a powerful and adaptive framework for automatic ECG-based arrhythmia diagnosis and holds strong potential for real-time clinical and wearable healthcare applications.