Patient-Aware Contrastive Learning Preserves Per-Patient Structure in RR-Interval Representations

Contrastive representation learning is widely used for physiological signals, but standard methods often struggle when individual subjects exhibit distinct baseline patterns. When class-level objectives (like supervised contrastive learning) are applied, they tend to collapse each subject's unique physiology into a single, shared class cluster. This loss of individual variation makes it difficult for models to generalize to unseen patients, as the learned embeddings lack the per-patient consistency required for reliable transfer.

The authors propose a patient-aware contrastive objective designed to solve this by explicitly incorporating subject identity into the training process. Instead of treating all same-class examples as positives, the model only forms positive pairs from segments that share both the same class label and the same patient identifier. This forces the encoder to learn representations that distinguish between classes while maintaining the specific geometric structure of each patient's sinus rhythm (SR) baseline.

By examining the learned embedding geometry, the researchers identified a "BCE paradox": while binary cross-entropy (BCE) produces the cleanest global class separation, it results in the most disordered per-patient structure, leading to poor performance on unseen patients. In contrast, the proposed patient-aware objective achieved the highest per-patient SR cohesion (0.850) compared to supervised contrastive learning (0.800) and BCE (0.772). This geometric consistency translates into superior downstream performance, reaching an AUROC of 0.989 on held-out patients with significantly lower seed-to-seed variance than standard baselines.

This work demonstrates that for physiological signals, global class separability is a misleading metric for model robustness. Instead, per-subject geometric consistency is the primary driver of cross-patient generalization. By ensuring that the model preserves individual patient characteristics, the proposed method provides a more stable and reliable framework for clinical applications like atrial fibrillation detection, where consistent performance across diverse patient populations is critical.