Bootstrap each lead's latent: A novel method for self-supervised learning of multilead electrocardiograms

Area of Science:

• Cardiology
• Artificial Intelligence
• Machine Learning

Background:

• Electrocardiograms (ECGs) are crucial for diagnosing cardiovascular diseases (CVDs).
• Deep learning models show promise for automated ECG analysis but require extensive manual labeling.
• Manual ECG labeling is time-consuming and costly, hindering model development.

Purpose of the Study:

• To introduce a novel self-supervised learning (SSL) method called Bootstrap Each Lead's Latent (BELL) for multilead ECG analysis.
• To reduce the dependency on manually labeled ECG data for training deep learning models.
• To enhance the performance of deep learning models in various downstream tasks, particularly with limited training data.

Main Methods:

• BELL is a variant of Bootstrap Your Own Latent (BYOL), adapted for multilead ECGs.
• It employs a multiple-branch architecture to process multilead ECG data effectively.
• Pretraining utilizes intra-lead and inter-lead mean squared error (MSE) loss functions without negative pairs, enhancing efficiency.

Main Results:

• BELL outperforms previous methods in experimental evaluations.
• Pretraining with BELL improved model performance by 0.69%–8.89% in downstream tasks using only 10% of available training data.
• The method demonstrated strong adaptability to uncurated, real-world hospital ECG data with minimal performance degradation (<1%).

Conclusions:

• BELL effectively alleviates the reliance on manual ECG labeling, addressing a key bottleneck in current deep learning applications.
• This approach facilitates the broader application of deep learning in automatic ECG analysis.
• BELL can help reduce the diagnostic burden on cardiologists in clinical settings.