acceptodds
Under review as a conference paper at ICLR 2027

PhysioDyT: Physiologically Directed Dynamic Transport for ECG–PPG Representation Learning

Abstract

Electrocardiography (ECG) and photoplethysmography (PPG) provide complementary views of cardiovascular dynamics: cardiac electrical activity precedes and drives peripheral hemodynamic responses, while the delay between the two modalities varies across individuals, time, and acquisition conditions. Existing multimodal representation learning methods commonly rely on synchronous alignment or symmetric similarity objectives, which may obscure this directed and time-varying relationship. We introduce PhysioDyT, a cross-modal predictive representation learning framework based on physiologically directed dynamic transport. Rather than enforcing fixed one-to-one correspondence, PhysioDyT predicts sample-dependent distributions over positive temporal offsets, preserves temporal ordering through soft monotone transport, and assigns unreliable or unmatched events to a dustbin state. The physiologically directed transport pathway is combined with bidirectional latent prediction to retain information useful to both sensing modalities. We evaluate PhysioDyT on a strictly patient-disjoint and pretraining-disjoint wearable cohort, including PPG-only coronary heart disease prediction across independent pretraining seeds. We further investigate the learned mechanism using direction reversal, cross-patient mismatch, temporal-shift interventions, and transport-component ablations. The experiments assess whether the model responds to physiological direction, patient-specific pairing, and temporal perturbations rather than relying solely on synchronous waveform similarity. Our results support dynamic transport as a structured inductive bias for cross-modal physiological representation learning, while also showing that the learned latent delay should not be interpreted as a direct clinical pulse-arrival-time measurement. PhysioDyT provides a principled and testable framework for learning representations from asynchronously coupled physiological signals.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Self-Supervised Learning, Multimodal Representation Learning, Physiological Time Series, ECG–PPG Modeling, Probabilistic Transport

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_WT1uOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes PhysioDyT, a JEPA-style ECG and PPG pretraining method, in which each ECG token is softly matched to PPG tokens one to five tokens ahead, through unbalanced optimal transport with a "dustbin" state. Subsequently only PPG is used for downstream. Using paired ECG for pretraining is very reasonable, and have of course, been used in a lot of sota work. However, this paper does not demonstrate much improvement in terms of downstream performance, when comapred with exsiting encoders.

Critical strengths
  • The comparison protocol is solid and reasonable.
  • The usage of ECG for pretraining on PPG is valid.
Critical weaknesses

I am mainly concerned that the paper does not show a clear downstream advantage, and that none of the proposed components is ablated.

Clarity and Presentation

  • the paper reads more like a progress report than a finished paper, e.g. "archived summary", "run-level verification is pending" (L435), and the test set "remains sealed" (L225).
  • Neither the pretraining corpus nor CWPC is described. Supplementary materials L390 said "will provide" configurations, so it is very hard to reproduce as submitted.
  • The related work could be benefited from more comprehensive and thorough discussion, currently on 10 references (!)

Methodology

  • None of the core components (positive delays, delay scoring, etc) is ablated.
  • I am not convinced by the physiologically directed claim. The learned delay barely agrees with pulse arrival time (Spearman ~0.022, MAE 256 ms), and a 160 ms token is far too coarse to resolve beat-to-beat PAT anyway.

Experiment and Results

  • I am a bit concerned about Table 1. It still reports the historical cohort (0.7695), and PhysioDyT drops by about 0.05 on the clean cohort, while MOMENT and NormWear barely change.
  • Pulse-PPG collapses to 0.5146 in the clean evaluation, what happened?
  • CHD detection from PPG might be strongly confounded by age, so an age/sex-only baseline would be expected.

Minor Issues and Suggestions

  • Figure could be enhanced by increasing the font size.
Questions for authors

Apart from above weakness, I would expect some clarification on:

  1. How does PhysioDyT compare with a PPG-only self-supervised model trained on the same corpus, and with variants without transport or without the dustbin?
  2. Could the authors describe the pretraining corpus and CWPC?
Reviewer_ZXQ8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies whether ECG–PPG timing can improve PPG representation learning, using paired signals only during training, which is a relevant setting. However, the central objective is unclear, the novelty is limited, the evaluation is incomplete, and the physiological interpretation is weakly supported.

Critical strengths

Investigating whether physiological timing between ECG and PPG can improve PPG representation learning is worthwhile. Using paired signals during training while requiring only PPG at deployment is a relevant research setting.

Critical weaknesses

Weaknesses

  1. The paper does not clearly distinguish its central objective: improving downstream prediction, learning physiological correspondence, or demonstrating sensitivity to temporal interventions. These objectives require different evidence, and their connection is insufficiently established.
  2. The technical innovation is limited, and the claimed contribution is not convincing. The method primarily combines established latent-prediction and alignment components with physiological constraints, without clearly establishing a substantial methodological advance beyond this integration.
  3. The experimental evaluation remains incomplete, with essential downstream ablations missing to establish the contribution of individual transport constraints. The main three-pre-training-seed result also awaits run-level verification, and the held-out test set has not been evaluated.
  4. The clean frozen-feature comparison does not reproduce the advantage reported on the historical development cohort. Moreover, stronger responses to token-space interventions than waveform shifts, together with weak agreement with waveform-derived pulse arrival time, leave the physiological interpretation insufficiently supported.
  5. The related-work discussion provides only a brief overview of representation learning and temporal alignment. It does not sufficiently distinguish the proposed method from closely related constrained alignment and cross-modal learning approaches, further weakening the positioning of its claimed novelty.
Questions for authors

Questions:

  1. Which is the central objective: downstream prediction, physiological correspondence, or sensitivity to temporal interventions?
  2. What does each transport constraint contribute downstream, and what are the held-out test-set results?
  3. Since the clean comparison does not reproduce the advantage and agreement with pulse arrival time is weak, what supports the physiological interpretation?
Reviewer_kpMvOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents an approach to ECG-PPG representation learning, leveraging the delay between ECG and PPG signals. The authors develop a JEPA style architecture with both modalities used during pretraining, but testing involves PPG alone. The pre-training is performed on a clinical dataset comprising both modalities whereas the evaluation is done on a clinical wearable dataset. The approach outperforms other baselines on the clinical dataset.

Critical strengths

The main strength of this paper is that it leverages the delay between ECG and PPG signals, as a source of self-supervision. This has not been utilized before for SSL.

Critical weaknesses
  • Overall, the writing is quite hard to follow with unclear sentences that are in many cases abrupt. The authors are urged to go through the entire paper and clear up the language. The paper would benefit greatly from it. Some examples include:

    • Sec 5.1 -> “Its run level patient predictions still require source level verification.” Not sure what this means.
    • Introduction: “Expected target positions are regularized for …. while the dustbin prevents low quality… from being forced into a match”
  • It is quite hard to understand the approach and even after reading through it a few times, I do not fully understand it. Concepts are suddenly referenced without being introduced first. The figure does not make it easy either.

    • For example, in Sec 3.2, the authors write: “For ECG token i and future offset d, the transport head combines a 128-dimensional projected content similarity c_i,i+d, global delay logits G_d from pooled ECG/PPG features, and local ECG-conditioned delay logits L_i,d”. Here, neither of these variables are introduced before or after. What are these delay logits G_d and L_i,d? The only place they are referred to is the figure.
    • The authors should do a substantial rewrite of their method, emphasizing clarity. There is ample space left (2 pages) so the approach can be expanded upon to make things more clear.
    • The figure is not easy to follow either, and needs substantial rework.
  • The pretraining and evaluation dataset need more details. It seems like the authors used a clinical dataset comprising both sensors. Is this a private dataset? What are the characteristics? Only the user counts in splits are provided but how many total samples are available? What are the demographics? How does it compare to public datasets such as MIMIC? What about IRB?

  • Going further, the authors should add comparison of PhysioDyT on publicly available PPG datasets. While I appreciate the public encoder baselines they have added, it is hard to get a feel for how well methods are expected to perform.

  • The related work section is rather sparse, and the paper contains only 10 references. Please expand this appropriately.

  • There are now newer and more effective PPG baselines, please compare against them. For example, here are a couple: [1,2].


References:

[1] Nie, Guangkun, Xiaocheng Fang, Gongzheng Tang, et al. “AnyPPG: An ECG-Guided PPG Foundation Model Trained on Over 100,000 Hours of Recordings for Holistic Health Profiling.” In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD ‘26). ACM, 2026. https://doi.org/10.1145/3770855.3818911.

[2] Gu, Xiao, Wei Tang, Jinpei Han, et al. “Cardiac Health Assessment Across Scenarios and Devices Using a Multimodal Foundation Model Pretrained on Data from 1.7 Million Individuals.” Nature Machine Intelligence 8, no. 2 (2026): 220–33. https://doi.org/10.1038/s42256-026-01180-5.

Questions for authors
  • For Sec 5.1, where is the validation performance coming from? I do not see any table containing that number. Further, what does the following text mean: “Its run-level patient predictions still require source-level verification. The 2,134-patient test partition remains sealed.” Could the authors clarify?
  • In the Appendix Section A, the authors mention that run-level verification is pending. Is this right?
  • Not clear what a historical-development cohort is. Please define it clearly.
  • How good is the learned ECG encoder on ECG datasets? Often, training with multiple modalities results in good encoders for all modalities. Is it the case for PhysioDyT as well?
  • Section 5.3 refers to Figure 3(c), and mentions reversal, negative delay cross-patient mismatch etc as counterfactual evaluation. However, the setting for these analyses is not clear. For example, what does reversal correspond to?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.