Joint spatio-temporal features constrained self-supervised electrocardiogram representation learning.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Cardiology

Background:

• Electrocardiogram (ECG) diagnostics are limited by reliance on labeled data.
• Existing self-supervised methods for ECG lack efficacy or require extensive expert annotation.
• There is a need for efficient, less supervised methods for ECG representation learning.

Purpose of the Study:

• To develop a self-supervised learning method for general ECG representations.
• To reduce dependence on labeled ECG data for downstream diagnostic tasks.
• To improve the efficiency and accuracy of ECG analysis.

Main Methods:

• Proposed a spatio-temporal joint detection self-supervised method for ECG.
• Dynamically masked ECG signals (temporal) and disrupted lead order (spatial).
• Reconstructed original signals and predicted lead numbers for model training.

Main Results:

• The method effectively learns ECG representations, validated on public and private datasets.
• Achieved similar performance to supervised methods using only 60% of labeled data.
• Demonstrated 1.3% and 8.6% accuracy improvement in classification and localization tasks compared to random initialization.

Conclusions:

• Self-supervised learning is feasible for learning effective ECG representations.
• The proposed method offers a promising alternative to traditional supervised learning for ECG analysis.
• This approach can significantly reduce the need for expert-annotated ECG datasets.