Clinical knowledge-based ECG abnormalities detection using dual-view CNN-Transformer and external attention mechanism

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Automatic Electrocardiogram (ECG) analysis is crucial for early cardiovascular disease detection and diagnosis.
• Deep neural networks are widely used but often suffer from poor interpretability and high complexity.
• Existing models struggle to achieve cardiologist-level performance in detecting ECG abnormalities.

Purpose of the Study:

• To develop a clinical knowledge-based ECG abnormalities detection model.
• To address limitations of traditional deep neural networks in ECG analysis.
• To improve interpretability and reduce complexity in automatic ECG abnormality detection.

Main Methods:

• Proposed a dual-view CNN-Transformer model incorporating an external attention mechanism.
• Mimicked clinician diagnosis by considering both detailed waveform and global recording changes.
• Integrated external attention mechanisms to focus on informative ECG regions.

Main Results:

• Achieved an average F1-score of 0.854±0.01 on a 9-class dataset.
• Demonstrated superior performance compared to state-of-the-art models.
• Showcased enhanced interpretability and reduced computational complexity.

Conclusions:

• The proposed model offers a credible and effective solution for automatic ECG abnormality detection.
• The dual-view CNN-Transformer with external attention shows promise for clinical applications.
• This approach advances computer-aided diagnosis in cardiovascular medicine.