Universal representations in cardiovascular ECG assessment: A self-supervised learning approach.

Area of Science:

• Cardiology
• Artificial Intelligence
• Machine Learning

Background:

• The 12-lead electrocardiogram (ECG) is crucial for cardiovascular assessment.
• Deep learning for ECG analysis is promising but limited by labeled data.
• Self-supervised learning (SSL) offers a solution by learning from unlabeled data.

Purpose of the Study:

• Develop and validate an SSL methodology for universal ECG representations.
• Utilize longitudinal ECG data for robust cardiovascular assessments.
• Improve ECG analysis across various cardiovascular conditions.

Main Methods:

• A contrastive SSL pre-trained model was developed using over 4.9 million ECG tracings.
• The model learned universal ECG representations from longitudinal patient data.
• Evaluated on internal and external datasets covering diverse cardiovascular conditions.

Main Results:

• The SSL model demonstrated performance comparable to supervised models for various arrhythmias and myocardial infarction.
• Learned ECG representations significantly enhanced classification models with small sample sizes (up to 0.3 AUROC improvement).
• SSL approach proved effective in improving model robustness and learning.

Conclusions:

• SSL-derived ECG representations from longitudinal data are highly effective.
• This method particularly benefits analyses with limited sample sizes.
• The approach enhances learning processes and boosts the robustness of ECG-based cardiovascular assessments.