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

In clinical electrocardiogram (ECG) analysis, high-quality annotations are expensive and difficult to scale, leaving many tasks in an extreme few-shot learning regime. We formulate a single-label Top-5 rhythm classification task on PTB-XL and strictly limit supervised training to N = 70 labeled samples (14 per class) to characterize failure modes of scratch training and assess the stabilizing effect of self-supervised learning (SSL). We first perform SimCLR-style contrastive pretraining with the NT-Xent loss on 16,304 unlabeled ECG recordings, followed by supervised fine-tuning. To isolate the independent contribution of SSL initialization, downstream augmentation is disabled (NoAug) in the core evaluation. Performance and stability are assessed using Macro-F1, best validation epoch, and a collapse rate defined as Best Epoch ≤ 1. Under N = 70, scratch training exhibits systematic early collapse (Macro-F1 = 0.115 ± 0.072; collapse rate = 66.7%), which is not alleviated by strong downstream augmentation alone (0.108 ± 0.042; 66.7%). In contrast, SSL fine-tuning without downstream augmentation improves Macro-F1 to 0.192 ± 0.003 and reduces the collapse rate to 0%, with markedly reduced inter-seed variance. A balanced N = 300 reference achieves Macro-F1 = 0.297 ± 0.018. These results indicate that, in extreme few-shot ECG classification, the primary bottleneck is training reproducibility rather than peak accuracy, and SSL contrastive pretraining provides a robust initialization that substantially mitigates collapse and improves usability.