pith. sign in

arxiv: 2606.23570 · v1 · pith:6VLETS7Gnew · submitted 2026-06-22 · 💻 cs.LG

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

Yasantha Niroshana , Weijith Wimalasiri , Chathuranga Hettiarachchi This is my paper

Pith reviewed 2026-06-26 08:54 UTC · model grok-4.3

classification 💻 cs.LG
keywords contrastive learningRR intervalsatrial fibrillationpatient-awarephysiological signalsrepresentation learninggeneralizationPAF detection
0
0 comments X

The pith

Patient-aware contrastive learning preserves per-patient structure for better generalization to unseen patients in AF detection from RR intervals.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Standard contrastive methods on physiological signals like RR intervals tend to merge individual patient baselines into class clusters, losing the variation needed for generalizing to new patients. The paper proposes forming positive pairs only from segments of the same patient and same class to keep each patient's sinus rhythm baseline intact while separating atrial fibrillation from normal rhythm. This results in more consistent per-patient structure in the learned embeddings and higher accuracy on patients not seen in training. The work shows that maintaining subject-specific geometry matters more for cross-patient performance than achieving the cleanest overall class separation.

Core claim

By forming positive pairs exclusively from same-patient, same-class RR-interval segments, the patient-aware contrastive objective preserves each patient's unique sinus rhythm baseline while separating the paroxysmal atrial fibrillation class, leading to per-patient SR cohesion of 0.850 and a patient-independent AUROC of 0.989 ± 0.003 on the IRIDIA-AF dataset.

What carries the argument

Patient-aware contrastive objective that restricts positive pairs to same-patient same-class segments to maintain per-patient geometric structure.

If this is right

  • Per-patient SR structure cohesion is highest at 0.850 compared to 0.800 for SupCon and 0.772 for BCE.
  • The representation achieves 0.989 AUROC with 2.6 times lower seed variance than supervised contrastive baselines.
  • Binary cross-entropy produces clean global class separation but the most disordered per-patient structure, causing poor performance on unseen patients.
  • Per-subject geometric consistency is more important than global class separability for robust cross-patient generalization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The method may extend to other time-series medical signals where inter-subject variability dominates class differences.
  • Alternative pair selection strategies could be explored to balance patient consistency with class discrimination in contrastive frameworks.
  • Validation on diverse datasets from multiple centers would test if the preserved structure holds across different recording conditions.

Load-bearing premise

Forming positive pairs only from same-patient same-class segments is enough to keep per-patient structure while learning features that work across different patients.

What would settle it

If experiments on new patient cohorts show that the patient-aware method does not improve per-patient cohesion or reduce AUROC variance compared to standard losses, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 2606.23570 by Chathuranga Hettiarachchi, Weijith Wimalasiri, Yasantha Niroshana.

Figure 1
Figure 1. Figure 1: Standard supervised contrastive learning (left) treats all same-class segments as positives regardless of subject. The proposed patient-aware formulation (right) restricts positives to same-patient, same-class segments. 2.2. Patient-Aware Mini-Batch Sampling Each mini-batch samples P patients, each contributing n SR and n AF windows (B = 2nP). A patient is eligible only if both its SR and AF pools contain … view at source ↗
Figure 2
Figure 2. Figure 2: Per-metric linear-probe comparison of the proposed patient-aware objective, SupCon, and BCE on the held-out patient￾independent test split (5 seeds). Bars show the seed mean and error bars show ± one standard deviation. to weight initialization, which is a desirable property for any deployment that must reproduce model behaviour. AUROC improvements of +0.006 over SupCon and +0.009 over BCE each exceed one … view at source ↗
Figure 3
Figure 3. Figure 3: Episode selection and preprocessing pipeline. SR normalization windows (hour 0–1) are disjoint from classification windows (hours 1–2), ensuring the RobustScaler has no access to labeled data. The ≥ 4 hr SR inclusion criterion further ensures that all SR classification windows begin at least 2 hr before AF onset, reducing the risk of including pre-episode transitional rhythms. Per-patient normalization use… view at source ↗
Figure 4
Figure 4. Figure 4: Encoder architecture: multi-branch 1D-CNN backbone with temporal attention pooling and a two-layer MLP projection head. Architecture. Three parallel 1D-CNN branches with kernel sizes k∈ {3, 5, 7} capture beat-to-beat fluctuations, medium￾range oscillations, and broader trend dynamics, motivated by Inception style multi-scale processing (Szegedy et al., 2015). Each branch applies three strided convolution b… view at source ↗
read the original abstract

Contrastive representation learning struggles on physiological signals when each subject contributes a distinct baseline pattern. If class differences overlap with subject differences,class-level objectives such as supervised contrastive learning tend to merge per-subject structure into a single per-class cluster,removing the individual variation that a model needs to generalize to unseen patients. We study this problem in the setting of Paroxysmal Atrial Fibrillation(PAF) detection from RR-interval(RRI) sequences and propose a patient-aware contrastive objective that forms positive pairs only from same-patient, same-class segments, preserving each patient's own sinus rhythm(SR) baseline while still pushing the two classes apart. Examining the learned embeddings directly, our objective achieves the most consistent per-patient SR structure (cohesion $0.850$ vs. $0.800$ for supervised contrastive loss (SupCon) and $0.772$ for binary cross-entropy (BCE)). We also identify that BCE produces the cleanest global class separation yet the most disordered per-patient structure. This is precisely why a linear probe trained on its features breaks down on unseen patients. On the IRIDIA-AF dataset, the resulting representation reaches a patient-independent Area Under the Receiver Operating Characteristic Curve (AUROC) of $0.989 \pm 0.003$ with $2.6\times$ lower seed variance than supervised contrastive baselines.These results highlight that per-subject geometric consistency, rather than global class separability, is key to robust cross-patient generalization.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a patient-aware contrastive objective for learning RR-interval representations in Paroxysmal Atrial Fibrillation detection. Positive pairs are formed exclusively from same-patient, same-class segments to preserve per-patient sinus rhythm baselines while separating classes. On the IRIDIA-AF dataset this yields embedding cohesion of 0.850 (vs. 0.800 for SupCon and 0.772 for BCE) and patient-independent AUROC of 0.989 ± 0.003 with 2.6× lower seed variance than supervised contrastive baselines.

Significance. If the results hold, the work provides concrete evidence that per-patient geometric consistency, rather than global class separability, drives robust cross-patient generalization on physiological signals. The direct empirical comparison of cohesion and downstream AUROC on held-out patients, together with the explicit test of the same-patient pair design choice, strengthens the mechanistic claim and may generalize to other subject-varying time-series tasks.

major comments (2)
  1. [Results] Results section: the cohesion scores (0.850 vs. 0.800 vs. 0.772) are central to the claim that the proposed loss best preserves per-patient SR structure, yet the manuscript provides no equation, pseudocode, or verification that this metric is computed identically across the three methods; any difference in normalization, sampling, or distance definition would undermine the comparison.
  2. [Method] Method section: the patient-aware contrastive loss is introduced without an explicit derivation showing how the same-patient positive-pair constraint modifies the gradient or embedding geometry relative to standard SupCon; this makes it difficult to assess whether the reported gains are due to the intended mechanism or to incidental changes in pair statistics.
minor comments (2)
  1. [Abstract] Abstract: the cohesion metric is reported as a point estimate while AUROC includes ±0.003 error bars; adding variability measures or statistical tests for cohesion would improve comparability.
  2. [Results] The manuscript does not report the number of patients or segments used for the cohesion calculation, which would help readers assess the reliability of the 0.850 figure.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for minor revision. We address each major comment below and will update the manuscript accordingly.

read point-by-point responses

  1. Referee: [Results] Results section: the cohesion scores (0.850 vs. 0.800 vs. 0.772) are central to the claim that the proposed loss best preserves per-patient SR structure, yet the manuscript provides no equation, pseudocode, or verification that this metric is computed identically across the three methods; any difference in normalization, sampling, or distance definition would undermine the comparison.

    Authors: We agree that the cohesion metric requires an explicit definition and verification of identical computation. The current manuscript reports the scores but omits the equation and pseudocode. In the revised manuscript we will add the mathematical definition of cohesion (including normalization and distance), pseudocode for its computation, and an explicit statement confirming that the identical procedure was applied to all three methods. This will be inserted in the Results section. revision: yes

  2. Referee: [Method] Method section: the patient-aware contrastive loss is introduced without an explicit derivation showing how the same-patient positive-pair constraint modifies the gradient or embedding geometry relative to standard SupCon; this makes it difficult to assess whether the reported gains are due to the intended mechanism or to incidental changes in pair statistics.

    Authors: We acknowledge that a derivation would strengthen the mechanistic explanation. The manuscript presents the loss formulation but does not derive its gradient differences. In the revised version we will add a concise derivation in the Method section that shows how restricting positive pairs to same-patient, same-class segments alters the gradient relative to standard SupCon and affects embedding geometry, thereby clarifying that the gains arise from the intended per-patient constraint rather than incidental pair statistics. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical results on held-out data

full rationale

The paper proposes a patient-aware contrastive loss that restricts positive pairs to same-patient same-class segments and then reports measured embedding cohesion (0.850) and patient-independent AUROC (0.989 ± 0.003) on the IRIDIA-AF dataset. These quantities are computed on held-out patients after training; they are not algebraically forced by the loss definition or by any self-citation chain. No equation equates a claimed performance metric to a fitted parameter, no uniqueness theorem is imported from prior author work, and the central claim (per-patient structure aids cross-patient generalization) is tested rather than presupposed. The derivation therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central modeling choice is the domain assumption that same-patient same-class segments constitute the correct positive pairs; no new physical entities or free parameters beyond standard contrastive-loss hyperparameters are introduced in the abstract.

free parameters (1)
  • contrastive temperature
    Standard hyperparameter in contrastive objectives; value not reported in abstract.
axioms (1)
  • domain assumption Positive pairs should be restricted to same-patient same-class segments to preserve per-patient structure.
    This rule is the load-bearing change that defines the proposed objective.

pith-pipeline@v0.9.1-grok · 5817 in / 1332 out tokens · 38971 ms · 2026-06-26T08:54:44.310194+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 24 canonical work pages · 1 internal anchor

  1. [1]

    Sci Data , publisher =

    Gilon, Cédric and Grégoire, Jean-Marie and Mathieu, Marianne and Carlier, Stéphane and Bersini, Hugues , month = oct, year =. Sci Data , publisher =. doi:10.1038/s41597-023-02621-1 , abstract =

    work page doi:10.1038/s41597-023-02621-1

  2. [2]

    Circulation , author =

    2023. Circulation , author =. 2024 , keywords =. doi:10.1161/CIR.0000000000001193 , abstract =

    work page doi:10.1161/cir.0000000000001193 2023

  3. [3]

    Achieving

    Chen, Wenjing and Zheng, Peirong and Bu, Yuxiang and Xu, Yuanning and Lai, Dakun , month = sep, year =. Achieving. Bioengineering , publisher =. doi:10.3390/bioengineering11090903 , abstract =

    work page doi:10.3390/bioengineering11090903

  4. [4]

    Gavidia, Marino and Zhu, Hongling and Montanari, Arthur N. and Fuentes, Jesús and Cheng, Cheng and Dubner, Sergio and Chames, Martin and Maison-Blanche, Pierre and Rahman, Md Moklesur and Sassi, Roberto and Badilini, Fabio and Jiang, Yinuo and Zhang, Shengjun and Zhang, Hai-Tao and Du, Hao and Teng, Basi and Yuan, Ye and Wan, Guohua and Tang, Zhouping and...

    work page doi:10.1016/j.patter.2024.100970 2024

  5. [5]

    , volume =

    2020 ESC Guidelines for the diagnosis and management of atrial fibrillation developed in collaboration with the European Association for Cardio-Thoracic Surgery (EACTS). , volume =. European Heart Journal , author =. 2020 , doi =. doi:10.1093/eurheartj/ehaa612 , abstract =

    work page doi:10.1093/eurheartj/ehaa612 2020

  6. [6]

    Archives of Cardiovascular Diseases , author =

    Role of the autonomic nervous system and premature atrial contractions in short-term paroxysmal atrial fibrillation forecasting: Insights from machine learning models , volume =. Archives of Cardiovascular Diseases , author =. 2022 , keywords =. doi:10.1016/j.acvd.2022.04.006 , abstract =

    work page doi:10.1016/j.acvd.2022.04.006 2022

  7. [7]

    Rev Cardiovasc Med , author =

    Autonomic Nervous System Activity before Atrial Fibrillation Onset as Assessed by Heart Rate Variability , volume =. Rev Cardiovasc Med , author =. 2025 , keywords =. doi:10.31083/RCM25364 , abstract =

    work page doi:10.31083/rcm25364 2025

  8. [8]

    Bootstrap

    Grill, Jean-Bastien and Strub, Florian and Altché, Florent and Tallec, Corentin and Richemond, Pierre and Buchatskaya, Elena and Doersch, Carl and Avila Pires, Bernardo and Guo, Zhaohan and Gheshlaghi Azar, Mohammad and Piot, Bilal and kavukcuoglu, koray and Munos, Remi and Valko, Michal , year =. Bootstrap. Advances in

  9. [9]

    Sci Rep , author =

    Temporal and spatial self supervised learning methods for electrocardiograms , volume =. Sci Rep , author =. 2025 , keywords =. doi:10.1038/s41598-025-90084-2 , abstract =

    work page doi:10.1038/s41598-025-90084-2 2025

  10. [10]

    and Healy, Graham and Smeaton, Alan F

    Contrastive. IEEE Access , author =. 2020 , keywords =. doi:10.1109/ACCESS.2020.3031549 , abstract =

    work page doi:10.1109/access.2020.3031549 2020

  11. [11]

    A simple framework for contrastive learning of visual representations , volume =

    Chen, Ting and Kornblith, Simon and Norouzi, Mohammad and Hinton, Geoffrey , month = jul, year =. A simple framework for contrastive learning of visual representations , volume =. Proceedings of the 37th

  12. [12]

    Supervised contrastive learning,

    Khosla, Prannay and Teterwak, Piotr and Wang, Chen and Sarna, Aaron and Tian, Yonglong and Isola, Phillip and Maschinot, Aaron and Liu, Ce and Krishnan, Dilip , month = mar, year =. Supervised. doi:10.48550/arXiv.2004.11362 , abstract =

    work page doi:10.48550/arxiv.2004.11362 2004

  13. [13]

    , month = jul, year =

    Kiyasseh, Dani and Zhu, Tingting and Clifton, David A. , month = jul, year =. Proceedings of the 38th

  14. [14]

    PLoS Comput Biol , author =

    Patient contrastive learning:. PLoS Comput Biol , author =. 2022 , keywords =. doi:10.1371/journal.pcbi.1009862 , abstract =

    work page doi:10.1371/journal.pcbi.1009862 2022

  15. [15]

    Europace , author =

    Temporal patterns and short-term progression of paroxysmal atrial fibrillation: data from. Europace , author =. 2020 , pages =. doi:10.1093/europace/euaa123 , abstract =

    work page doi:10.1093/europace/euaa123 2020

  16. [16]

    and Shah, Amit J

    Shashikumar, Supreeth P. and Shah, Amit J. and Clifford, Gari D. and Nemati, Shamim , month = jul, year =. Detection of. Proceedings of the 24th. doi:10.1145/3219819.3219912 , abstract =

    work page doi:10.1145/3219819.3219912

  17. [17]

    Going deeper with convolutions , issn =

    Szegedy, Christian and Liu, Wei and Jia, Yangqing and Sermanet, Pierre and Reed, Scott and Anguelov, Dragomir and Erhan, Dumitru and Vanhoucke, Vincent and Rabinovich, Andrew , month = jun, year =. Going deeper with convolutions , issn =. 2015. doi:10.1109/CVPR.2015.7298594 , abstract =

    work page doi:10.1109/cvpr.2015.7298594 2015

  18. [18]

    Proceedings of the European Conference on Computer Vision (ECCV) , month =

    Wu, Yuxin and He, Kaiming , title =. Proceedings of the European Conference on Computer Vision (ECCV) , month =

  19. [19]

    Understanding

    Wang, Tongzhou and Isola, Phillip , month = nov, year =. Understanding. Proceedings of the 37th

  20. [20]

    J Electrocardiol , author =

    An automated detection of atrial fibrillation from single-lead. J Electrocardiol , author =. 2022 , keywords =. doi:10.1016/j.jelectrocard.2022.07.069 , abstract =

    work page doi:10.1016/j.jelectrocard.2022.07.069 2022

  21. [21]

    Front Physiol , author =

    An. Front Physiol , author =. 2021 , pages =. doi:10.3389/fphys.2021.657304 , abstract =

    work page doi:10.3389/fphys.2021.657304 2021

  22. [22]

    Saiz-Vivo, Javier and Corino, Valentina D. A. and Hatala, Robert and de Melis, Mirko and Mainardi, Luca T. , month = may, year =. Heart. Front. Physiol. , publisher =. doi:10.3389/fphys.2021.672896 , abstract =

    work page doi:10.3389/fphys.2021.672896 2021

  23. [23]

    Artif Intell Med , author =

    Complexity and spectral analysis of the heart rate variability dynamics for distant prediction of paroxysmal atrial fibrillation with artificial intelligence methods , volume =. Artif Intell Med , author =. 2008 , keywords =. doi:10.1016/j.artmed.2008.03.009 , abstract =

    work page doi:10.1016/j.artmed.2008.03.009 2008

  24. [24]

    Comput Biol Med , author =

    Automated detection of atrial fibrillation using long short-term memory network with. Comput Biol Med , author =. 2018 , keywords =. doi:10.1016/j.compbiomed.2018.07.001 , abstract =

    work page doi:10.1016/j.compbiomed.2018.07.001 2018

  25. [25]

    Expert Systems with Applications , author =

    A deep learning approach for real-time detection of atrial fibrillation , volume =. Expert Systems with Applications , author =. 2019 , keywords =. doi:10.1016/j.eswa.2018.08.011 , abstract =

    work page doi:10.1016/j.eswa.2018.08.011 2019

  26. [26]

    Information Sciences , author =

    Dense lead contrast for self-supervised representation learning of multilead electrocardiograms , volume =. Information Sciences , author =. 2023 , keywords =. doi:10.1016/j.ins.2023.03.099 , abstract =

    work page doi:10.1016/j.ins.2023.03.099 2023

  27. [27]

    Computers in Biology and Medicine , author =

    A novel multimodal self-supervised framework for. Computers in Biology and Medicine , author =. 2025 , keywords =. doi:10.1016/j.compbiomed.2025.111137 , abstract =

    work page doi:10.1016/j.compbiomed.2025.111137 2025

  28. [28]

    Enhancing

    Sun, Xiaoyu and Yang, Yang and Dong, Xunde , month = may, year =. Enhancing. doi:10.48550/arXiv.2506.06310 , abstract =

    work page doi:10.48550/arxiv.2506.06310

  29. [29]

    Akiba, Takuya and Sano, Shotaro and Yanase, Toshihiko and Ohta, Takeru and Koyama, Masanori , month = jul, year =. Optuna:. doi:10.48550/arXiv.1907.10902 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1907.10902 1907