Arrhythmia classification based on multi-input convolutional neural network with attention mechanism.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Arrhythmia poses significant health risks, including stroke and cardiac arrest.
• Deep learning has improved ECG analysis, but challenges like signal variability and data imbalance persist.
• Existing methods often struggle with feature representation and unimodal inputs.

Purpose of the Study:

• To develop a novel arrhythmia classification algorithm using a multi-input convolutional neural network (CNN) with a Squeeze-Excitation (SE) attention mechanism.
• To integrate multi-scale time-frequency representations from ECG signals for enhanced feature learning.
• To improve the accuracy and robustness of automated arrhythmia detection.

Main Methods:

• A dual-branch CNN architecture processing ECG signals segmented into two temporal resolutions.
• Integration of multi-scale time-frequency representations using Short-Time Fourier Transform (STFT).
• Squeeze-Excitation (SE) blocks to enhance inter-channel dependencies and feature prioritization.
• A fusion strategy combining feature maps via bicubic interpolation and element-wise summation.

Main Results:

• The proposed model achieved high classification accuracy: 99.13% on the MIT-BIH database and 95.84% on the SPH database.
• Excellent Macro-F1 scores were obtained: 94.46% (MIT-BIH) and 95.91% (SPH).
• The model outperformed several existing state-of-the-art arrhythmia classification methods.

Conclusions:

• The novel multi-input CNN with SE attention demonstrates superior performance in arrhythmia classification.
• The integration of multi-scale features and attention mechanisms enhances robustness and accuracy.
• The proposed algorithm shows significant potential for clinical application in arrhythmia diagnosis.