Patient contrastive learning: A performant, expressive, and practical approach to electrocardiogram modeling

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Healthcare
• Cardiovascular Digital Health

Background:

• Supervised machine learning in healthcare faces challenges due to limited labeled training data.
• Developing robust models for clinical tasks requires effective feature extraction from medical signals like ECGs.

Purpose of the Study:

• To introduce Patient Contrastive Learning of Representations (PCLR), a novel pre-training method for ECGs.
• To improve the performance of machine learning models on diverse clinical tasks by leveraging large unlabeled ECG datasets.
• To provide a practical and performant method for generating expressive latent ECG representations.

Main Methods:

• Utilized contrastive learning on over 3.2 million unlabeled 12-lead ECGs to create patient-specific latent representations.
• Trained linear models on PCLR-generated representations and compared performance against training neural networks from scratch.
• Evaluated PCLR against other pre-training methods including supervised, autoencoder-based unsupervised, and multi-segment contrastive approaches.

Main Results:

• Training linear models on PCLR representations yielded an average 51% performance increase across six training set sizes and four clinical tasks (sex classification, age regression, left ventricular hypertrophy, atrial fibrillation).
• PCLR demonstrated significant performance benefits in three out of four tasks compared to other ECG pre-training methods.
• Achieved an average performance benefit of 47% over alternative models and a 9% benefit over the best-performing model for each task.

Conclusions:

• PCLR offers a powerful pre-training strategy to overcome data scarcity in healthcare machine learning.
• The generated ECG representations are expressive, performant, and broadly applicable to various clinical applications.
• PCLR provides a valuable tool for researchers and clinicians to enhance predictive modeling using ECG data.