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

Technical Summary: Patient-Aware Contrastive Learning for RR-Interval Representations

Problem Statement
Contrastive representation learning often struggles with physiological signals because individual subjects possess distinct baseline patterns. In standard formulations, such as supervised contrastive learning (SupCon), the assumption that all same-class examples are equally suitable as positive pairs breaks down when class differences overlap with subject differences. Consequently, class-level objectives tend to merge per-subject structures into a single per-class cluster. While this creates well-separated global classes, it removes the individual variation necessary for a model to generalize to unseen patients. The authors identify that a linear probe trained on such embeddings fails on new patients because the embedding space lacks consistent per-subject geometric structure, even if global class separability is high.

Methodology
The study focuses on Paroxysmal Atrial Fibrillation (PAF) detection using RR-interval (RRI) sequences from the IRIDIA-AF dataset. The authors propose a Patient-Aware Contrastive Objective designed to preserve per-patient structure while maintaining class separation.

• Positive Pair Construction: Unlike standard supervised contrastive learning, which treats all same-class segments as positives regardless of the subject, the proposed objective restricts positive pairs to segments that are both same-patient and same-class. Negative pairs are defined as segments from different classes.
• Sampling Strategy: The mini-batch sampling ensures that every patient contributes an equal number of Sinus Rhythm (SR) and Atrial Fibrillation (AF) windows, guaranteeing intra-patient, intra-class positive pairs in every batch.
• Architecture: The method utilizes a lightweight multi-branch 1D-CNN backbone (with kernel sizes 3, 5, and 7) followed by temporal attention pooling and a projection head. The contrastive loss is applied to the projected embeddings, while a linear probe is trained on the pooled outputs for downstream evaluation.
• Training Protocol: Experiments employ a frozen-encoder linear probing setup. The proposed method is compared against standard Supervised Contrastive Learning (SupCon) and Binary Cross-Entropy (BCE) baselines, with identical encoders, samplers, and probes to isolate the effect of the loss function.

Key Contributions

1. Patient-Aware Objective: A novel contrastive loss that forms positive pairs only from same-patient, same-class segments. This preserves each subject's individual SR baseline while still pushing the two classes (SR and AF) apart.
2. Embedding Geometry Analysis: The paper provides a direct analysis of the learned embedding space, revealing a critical trade-off:
   • BCE produces the cleanest global class separation (highest centroid distance) but results in the most disordered per-patient structure.
   • SupCon achieves high AF compactness but collapses distinct individual SR baselines into a shared prototype, reducing SR cohesion.
   • Proposed Method achieves the highest per-patient SR cohesion, maintaining the geometric consistency required for cross-patient transfer.
3. Downstream Validation: The method is validated on the IRIDIA-AF dataset for PAF detection, demonstrating superior performance and stability compared to existing baselines.

Results
On the patient-independent test split (held-out patients), the proposed method achieved the following:

• AUROC: 0.989 ± 0.003, outperforming SupCon (0.983 ± 0.009) and BCE (0.980 ± 0.012).
• Stability: The seed-to-seed variance in AUROC was reduced by 2.6× compared to SupCon and 3.4× compared to BCE, indicating a more robust and reproducible representation.
• Geometric Metrics: The proposed method achieved the highest SR cohesion (0.850) compared to SupCon (0.800) and BCE (0.772).
• Clinical Relevance: The method achieved an AF recall of 0.974 ± 0.008, exceeding the patient-holdout sensitivity of prior work (Andersen et al., 2019) by over 11 percentage points, despite using only a frozen encoder and a logistic regression probe.

Significance and Claims
The paper argues that per-subject geometric consistency, rather than global class separability, is the key determinant of robust cross-patient generalization in physiological signal analysis. The "BCE paradox" identified in the study demonstrates that high global class separation can coexist with poor transfer performance if the per-patient structure is disorganized. By restricting positive pairs to same-patient, same-class segments, the proposed objective prevents the collapse of individual baselines into a shared prototype.

The authors claim that this approach offers a broadly useful primitive for representation learning on signals organized by subject (e.g., EEG, EMG, PPG), provided that class labels and subject identifiers are available during training. The method does not depend on specific encoder architectures and is designed to be compatible with edge wearable hardware due to its lightweight backbone. The study concludes that preserving the individual structure of each subject is essential for models to extrapolate effectively to unseen patients.