Self-supervised pre-training with joint-embedding predictive architecture boosts ECG classification performance.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Cardiology

Background:

• Accurate heart arrhythmia diagnosis relies on electrocardiogram (ECG) interpretation.
• Automating ECG analysis is hindered by the scarcity and cost of large annotated datasets.
• Transfer learning and self-supervised learning (SSL) are key to overcoming data limitations in ECG classification.

Purpose of the Study:

• To investigate the efficacy of the Joint-Embedding Predictive Architecture (JEPA) for self-supervised learning on ECG data.
• To evaluate JEPA's performance against established SSL methods in ECG representation learning.
• To demonstrate JEPA's capability in pre-training models for downstream ECG classification tasks.

Main Methods:

• Utilized a large unsupervised dataset combining ten public ECG databases (>1 million records).
• Employed Vision Transformers pre-trained with JEPA, a non-generative, non-invariance-based SSL approach.
• Fine-tuned pre-trained models on PTB-XL benchmarks for arrhythmia classification.

Main Results:

• JEPA-pretrained models achieved superior performance compared to invariance-based and generative SSL methods.
• Attained an Area Under the Curve (AUC) of 0.945 on the PTB-XL 'all statements' task.
• Demonstrated consistent learning of high-quality representations, beneficial even with limited additional data.

Conclusions:

• JEPA offers a powerful alternative for self-supervised pre-training in ECG analysis, outperforming existing methods.
• JEPA's ability to learn robust representations without manual data augmentation or generative reconstruction is a significant advantage.
• This approach holds promise for improving automated ECG interpretation and diagnosis, especially in data-scarce scenarios.