A New Deep Learning Method with Self-Supervised Learning for Delineation of the Electrocardiogram.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Cardiology

Background:

• Electrocardiogram (ECG) characteristic points are crucial for cardiac diagnosis.
• Accurate detection of P-wave, QRS complex, and T-wave features is essential.
• Current methods may require extensive manual annotation.

Purpose of the Study:

• To propose a self-supervised deep learning framework for detecting ECG characteristic points.
• To utilize a modified Densenet architecture for enhanced feature extraction.
• To reduce reliance on human-annotated labels for ECG analysis.

Main Methods:

• Employed a self-supervised learning framework with a modified Densenet model.
• Pre-trained the model on large ECG datasets (QTDB, MITDB, NSRDB) using a pretext task of signal transformation discrimination.
• Transferred learned weights to a downstream task for characteristic point localization on the QT dataset.

Main Results:

• Achieved accurate detection of P-wave, QRS complex, and T-wave onset, peak, and termination points.
• Reported mean ± standard deviation detection errors for various characteristic points (e.g., P-wave termination: -0.28 ± 10.19).
• Demonstrated the efficacy of the self-supervised approach in ECG feature point detection.

Conclusions:

• The proposed self-supervised deep learning framework accurately detects crucial heartbeat feature points.
• This method provides a foundation for automated extraction of key information in ECG-based cardiac diagnosis.
• Self-supervised learning offers a promising avenue for ECG analysis without extensive manual labeling.