3DECG-Net: ECG fusion network for multi-label cardiac arrhythmia detection

Area of Science:

• Cardiology and Artificial Intelligence
• Medical Signal Processing

Background:

• Cardiovascular diseases are the leading cause of death globally.
• Electrocardiograms (ECG) are crucial for diagnosing heart conditions.
• Increased ECG data from wearables necessitates automated analysis.

Purpose of the Study:

• To introduce 3DECG-Net, a deep learning model for multi-label ECG analysis.
• To detect and classify seven distinct heart states using 12-lead ECG data.
• To develop an efficient ECG preprocessing framework.

Main Methods:

• Utilized a residual architecture with a multi-head attention mechanism.
• Transformed 12-lead ECG signals into 3D data using the Recurrent Plot technique.
• Employed a five-fold cross-validation scheme for model training.
• Applied Local Interpretable Model-agnostic Explanations (LIME) for AI explainability.

Main Results:

• Achieved a micro F1-score of 80.3%, outperforming state-of-the-art models.
• Demonstrated the model's ability to accurately diagnose specific arrhythmias.
• Generated compact, high-quality ECG signals through a novel preprocessing framework.

Conclusions:

• 3DECG-Net is a trustworthy and effective tool for diagnosing heart arrhythmias.
• The model can significantly improve diagnostic efficiency and facilitate early treatment.
• The findings support the potential clinical application of AI in cardiovascular care.