acceptodds
Withdrawn from ICLR 2027

Relational Window Stabilization for Repeatable Single-Window PPG Disease-Background Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Deep learning models for physiological signals are increasingly used for continuous monitoring of chronic disease background, yet their evaluation mainly relies on population-level discrimination metrics such as AUROC and AUPRC. In continuous monitoring, however, repeated measurements from the same individual should produce repeatable predictions when the underlying disease-background state remains unchanged over short intervals. Recent studies have revealed substantial prediction variability across repeated physiological measurements, highlighting a gap between discrimination performance and prediction repeatability. Here, we introduce Relational Window Stabilization, a training framework that learns repeatable single-window PPG disease-background predictions by exploiting relationships among multiple windows collected within the same monitoring episode. RWS uses multiple windows only during training, preserving individual-window supervision while reducing within-episode prediction variation without requiring inference-time aggregation. We evaluate RWS on the MC-MED PPG benchmark across nine chronic ICD-derived disease-background tasks using patient-disjoint chronological splits. Across different input durations, architectures, representations, and random seeds, RWS improves both discrimination and repeatability, increasing AUROC, AUPRC, and intraclass correlation coefficient while reducing prediction variability and diagnostic flip rates. These improvements remain under matched multi-window exposure controls and alternative repeatability-learning strategies. Cross-dataset evaluation on MIMIC-III further demonstrates that the repeatability advantage persists after checkpoint transfer and independent in-domain retraining. Together, these findings establish prediction repeatability as a complementary evaluation dimension for continuous physiological AI and demonstrate that same-episode relational supervision can improve single-window PPG predictions without changing the deployment paradigm

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Prediction Repeatability, Relational Learning, Physiological Signal Modeling, Deep Learning, Photoplethysmography, Continuous Monitoring

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_eRcfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes Relational Window Stabilization (RWS), a training-only framework for single-window PPG prediction of chronic disease background. During training, RWS uses K windows from the same episode. It adds three terms to per-window BCE: a leave-one-out logit consistency term, a logit-range hinge, and an ICC surrogate. Inference uses a single window. On MC-MED (9 ICD-derived tasks, 7 architectures, 3 durations) and on MIMIC-III-Ext-PPG, RWS is reported to improve AUROC/AUPRC and repeatability metrics (ICC, within-episode SD, flip rate, VDFR).

Critical strengths
  1. Prediction repeatability under repeated observations is a practically relevant and under-evaluated property for continuous monitoring.
  2. The evaluation is broad: multiple architectures and frozen encoders, a fair-exposure control (BCE-only K=8), alternative consistency methods, ablations, patient-disjoint splits, and cross-dataset transfer and retraining.
  3. Inference is unchanged (single window), so deployment cost does not increase.
Critical weaknesses

1 It is difficult to understand the key hyperparameters and how the baseline experiments were constructed. 2. Since multiple windows are readily available in continuous monitoring, the main text should include a comparison with a Conventional model that averages predictions over multiple windows at inference.

Questions for authors
  1. Consistency regularization across multiple views of the same instance is a well-established approach (e.g., mean teacher, R-Drop). Is there a clear distinction between RWS and these methods? (e.g., focus of cocept)
  2. Could you show the prediction distributions of Conventional and RWS? How can a within-episode SD of 0.001 coexist with a flip rate of 0.687 (e.g., PaPaGei 10 s)?
  3. How does RWS (D=1) compare with Conventional (D=4/8)?
  4. What is the motivation for using the leave-one-out target and stopgrad?
  5. Does RWS reduce sensitivity to genuine changes in physiological state within an episode?
Reviewer_HEmtOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper "Relational Window Stabilization for Repeatable Single-Window PPG Disease-Background Prediction" makes the case for improving the performance of PPG-based prediction models by leveraging multiple prediction windows. The authors consider the prediction of chronic conditions and therefore assume that all prediction windows carry the same label and employ different consistency losses to enforce consistent classification behavior. The authors demonstrate improved performance compared to single-window prediction models and simpler baselines.

Critical strengths
  • The idea of increasing the temporal context of prediction models based on physiological time series is very promising.
  • Code is provided in a well-structured code repository, which supports the reproducibility of the study.
  • The authors demonstrate improved performance compared to single-window prediction models and simpler baselines.
  • The authors consider two datasets and several models per dataset including foundation models.
Critical weaknesses

Major

  • "More recently, AI-based ECG studies have directly quantified prediction variability across repeated recordings and demonstrated that model outputs can change over short intervals even when the underlying clinical state is expected to remain unchanged" I fundamentally disagree with this statement. Unless the patient shows persistent AF, it is quite likely to have short AF episodes interleaved with SR.
  • Related to that: Chronic conditions are not that well-aligned with PPG as input modality, i.e., there are much better diagnostic means to detect the considered chronic conditions. Atrial fibrillation is not a good example as a patient is not necessarily in persistent AF. Unfortunately, the treatment becomes more complicated in the case where not all windows are assumed to show the same label.
  • Model insights remain limited to impact on global performance metrics and provide no further insights into the respective models.
  • The limitations are discussed only rather superficially in a short final discussion paragraph.

Minor

  • Introduction: "Recent clinical AI studies have provided direct evidence of this phenomenon." I would like to point to research in the sleep staging literature, where the importance of long temporal context has been recognized a long time ago (see e.g. https://arxiv.org/abs/1809.10932) and is widely accepted in corresponding prediction models. For AF detection this was recognized only very recently, see e.g. https://arxiv.org/abs/2510.17406. In particular the large body of literature work on the inclusion of temporal context in the domain of sleep staging should be properly acknowledged.
  • Section 3.1: The considered tasks along with precise definition are hard to parse from main text and appendix. A detailed description and a summarizing table would be very much appreciated.

Summary

  • This submission is a good fit for a biomedical journal, but has questionable fit to ICLR (in particular to the subdomain of unsupervised, self-supervised, semi-supervised, and supervised representation learning selected as primary area) as it presents very little insights beyond impact on global performance metrics.
  • I am fully supportive of the idea to exploit more temporal context, but I find the chronic disease prediction misaligned with PPG as input modality.
Questions for authors
  • How is ECGFounder pretrained on ECG adapted to the PPG domain?
Reviewer_VajfOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Relational Window Stabilization (RWS) to improve prediction repeatability for PPG-based chronic disease classification. The method combines conventional supervised classification with three additional regularization terms that encourage consistent predictions across windows from the same monitoring episode. Experiments across multiple architectures, input durations, and datasets show improved prediction repeatability and modest gains in discrimination performance.

Critical strengths

The motivation of improving prediction repeatability in continuous physiological monitoring is reasonable. The experiments also cover multiple architectures and include cross-dataset evaluation.

Critical weaknesses
  1. Limited methodological novelty. The proposed method mainly combines existing regularization techniques, including prediction consistency, logit-range constraints, and variance-based regularization. These components have substantially overlapping objectives, making the overall contribution rather incremental.

  2. Unnecessarily complicated loss design. The final objective combines four loss terms with multiple weighting coefficients. It is unclear whether all three relational losses are necessary or whether a simpler consistency regularizer could achieve comparable results. Additional sensitivity analyses of the loss coefficients would help justify the design.

  3. Limited practical significance of repeatability improvements. Since the proposed losses explicitly penalize within-episode prediction variation, improvements in repeatability metrics are largely expected. However, more consistent predictions do not necessarily indicate better clinical reliability, and excessive regularization may suppress meaningful physiological variations.

  4. Modest discrimination improvements. Despite the complexity of the proposed method, improvements in AUROC and AUPRC are relatively limited. In some settings, simpler alternatives achieve comparable or better discrimination performance, raising questions about whether the additional complexity is justified.

Questions for authors

N/A

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.