Stabilizing Extreme Few-Shot ECG Classification via Self-Supervised Contrastive Pretraining.

Area of Science:

• Cardiology
• Machine Learning
• Medical Informatics

Background:

• High-quality electrocardiogram (ECG) annotations are scarce, posing challenges for machine learning models in few-shot learning scenarios.
• Existing methods struggle with training stability and reproducibility when labeled data is extremely limited.

Purpose of the Study:

• To evaluate the effectiveness of self-supervised learning (SSL) for improving ECG rhythm classification under extreme few-shot conditions.
• To characterize failure modes of standard training and assess the stabilizing impact of SSL pretraining.

Main Methods:

• Contrastive pretraining using SimCLR with NT-Xent loss on a large unlabeled ECG dataset (16,304 recordings).
• Supervised fine-tuning on a small labeled subset (N=70) for Top-5 rhythm classification.
• Evaluation focused on training stability (collapse rate) and performance (Macro-F1), with and without downstream augmentation.

Main Results:

• Scratch training without SSL exhibited significant early collapse (66.7% rate) and low performance (Macro-F1=0.115).
• Downstream augmentation alone did not alleviate collapse or improve performance.
• SSL pretraining followed by fine-tuning drastically reduced the collapse rate to 0% and improved Macro-F1 to 0.192, with reduced variance.

Conclusions:

• In extreme few-shot ECG classification, training reproducibility is a critical bottleneck, more so than peak accuracy.
• SSL contrastive pretraining offers robust initialization, substantially mitigating training collapse and enhancing usability for limited-data ECG analysis.