Abstract
Arrhythmia is a prevalent cardiac disorder that can lead to severe complications such as stroke and cardiac arrest. While deep learning has advanced automated ECG analysis, challenges remain in accurately classifying arrhythmias due to signal variability, data imbalance, and feature representation limitations. In this work, we propose a novel arrhythmia classification algorithm based on a multi-input convolutional neural network (CNN) enhanced with a Squeeze-and-Excitation (SE) attention mechanism. Distinct from previous methods that rely on single-resolution features or unimodal inputs, our model integrates multi-scale time-frequency representations derived from Short-Time Fourier Transform (STFT) applied to ECG signals segmented into two temporal resolutions. The dual-branch CNN architecture enables complementary feature learning from both short and long segments, while SE blocks enhance inter-channel dependencies to prioritize critical features. The fusion strategy combines feature maps via bicubic interpolation and element-wise summation to maintain spatial integrity. Evaluated on MIT-BIH and SPH arrhythmia databases, the proposed model achieves high accuracy (99.13% and 95.84%, respectively) and Macro-F1 scores (94.46% and 95.91%), outperforming several state-of-the-art approaches. These results highlight the model's potential for robust and interpretable arrhythmia classification in clinical practice.