Diagnostic Feature Reconstruction for Enhanced Single-Lead ECG Classification.

Area of Science:

• Biomedical Engineering
• Cardiology
• Artificial Intelligence in Medicine

Background:

• Standard 12-lead electrocardiograms (ECG) are crucial for cardiovascular diagnosis but require clinical settings.
• Wearable few-lead ECG devices offer convenience but suffer from reduced diagnostic capability due to limited lead coverage.

Purpose of the Study:

• To enhance the diagnostic performance of single-lead ECGs by proposing a feature-reconstruction-based classification method.
• To bridge the informational gap between limited single-lead data and comprehensive 12-lead ECG diagnostics.

Main Methods:

• Utilized a pre-trained 12-lead ECG model to guide feature learning for single-lead signals.
• Employed a CNN-Transformer for multi-scale feature extraction and a transformer encoder for feature reconstruction.
• Integrated reconstructed and original single-lead features using cross-attention for enhanced classification.

Main Results:

• The proposed method effectively enhances feature discriminability for single-lead ECGs.
• Demonstrated improved single-lead ECG classification performance on two public datasets.
• Feature reconstruction, rather than signal reconstruction, avoids performance degradation.

Conclusions:

• The feature-reconstruction method significantly boosts single-lead ECG classification accuracy.
• Confirms the robustness and practical potential of the approach for wearable cardiovascular monitoring.
• Offers a viable solution to improve diagnostic capabilities of limited-lead ECG devices.