The article addresses the challenge of contrastive representation learning on physiological signals where subject-specific baselines interfere with class-level objectives, causing models to lose individual variation necessary for generalization. The authors propose a patient-aware contrastive objective for Paroxysmal Atrial Fibrillation detection that forms positive pairs only from same-patient segments to preserve sinus rhythm baselines while separating classes.
- The proposed method achieves the most consistent per-patient sinus rhythm structure with a cohesion score of 0.850, compared to 0.800 for supervised contrastive loss and 0.772 for binary cross-entropy.
- Binary cross-entropy produces the cleanest global class separation but results in the most disordered per-patient structure, causing linear probes to fail on unseen patients.
- On the IRIDIA-AF dataset, the representation achieves a patient-independent AUROC of 0.989 ± 0.003 with 2.6 times lower seed variance than supervised contrastive baselines.
The results indicate that maintaining per-subject geometric consistency is more critical for robust cross-patient generalization than achieving global class separability.