Temporal and spatial self supervised learning methods for electrocardiograms.

Area of Science:

• Biomedical Signal Processing
• Machine Learning
• Cardiology

Background:

• Supervised deep learning for electrocardiogram (ECG) detection is hindered by limited labeled data.
• Existing self-supervised learning methods for ECG analysis are often image-based, limiting their effectiveness.
• A need exists for advanced methods to improve ECG feature representation and detection accuracy.

Purpose of the Study:

• To introduce a novel Temporal-Spatial Self-Supervised Learning (TSSL) method for ECG detection.
• To leverage intrinsic temporal and spatial characteristics of ECG signals for enhanced feature representation.
• To overcome limitations of existing self-supervised learning approaches in ECG analysis.

Main Methods:

• Developed a Temporal-Spatial Self-Supervised Learning (TSSL) framework tailored for ECG signals.
• Exploited temporal ECG signal properties for stable individual representations across time.
• Utilized spatial ECG signal correlations across different leads to capture holistic cardiac activity.

Main Results:

• TSSL demonstrated superior performance on CPSC2018, Chapman, and PTB-XL databases compared to existing methods.
• The method achieved performance comparable to full-label training using only 10% of labeled data.
• TSSL provided enhanced feature extraction and deeper insights into ECG signal characteristics.

Conclusions:

• TSSL effectively utilizes temporal and spatial information in ECG signals for improved detection.
• The proposed method significantly enhances feature representation in self-supervised ECG analysis.
• TSSL offers a promising solution for ECG detection with limited labeled data, advancing the field.