Self-supervised representation learning from 12-lead ECG data.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Clinical 12-lead electrocardiography (ECG) is a crucial biosignal, but label scarcity hinders model training.
• Self-supervised learning (SSL) offers a solution to leverage large unlabeled ECG datasets.

Purpose of the Study:

• To comprehensively assess SSL for representation learning in clinical 12-lead ECG data.
• To evaluate the impact of SSL pretraining on downstream ECG classification tasks.

Main Methods:

• Adapted state-of-the-art SSL methods (instance discrimination, latent forecasting) to ECG data.
• Evaluated learned representations using linear evaluation and fine-tuning on a clinical ECG classification task.
• Utilized publicly available ECG datasets for comprehensive assessment.

Main Results:

• Contrastive predictive coding adaptation achieved linear evaluation performance close to supervised methods (-0.5%).
• Fine-tuned models showed ~1% improvement in downstream performance, enhanced label efficiency, and robustness to noise.
• SSL significantly improved ECG classifier performance compared to purely supervised training.

Conclusions:

• SSL is feasible for extracting discriminative representations from ECG data.
• SSL pretraining offers substantial advantages for downstream tasks over purely supervised training.
• This work provides a foundation for reproducible progress in ECG representation learning.