acceptodds
Withdrawn from ICLR 2027

PhysPPG: Physics-Informed Contrastive Learning for Photoplethysmography Representation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Photoplethysmography (PPG) enables accessible cardiovascular monitoring, yet motion and acquisition artifacts can obscure the waveform structure needed for reliable physiological inference. We introduce PhysPPG, a physics-informed contrastive learning framework for robust PPG representation learning. A temporal encoder maps PPG signals to latent trajectories, which are jointly shaped by a contrastive objective and a generalized Euler–Lagrange residual. Learnable kinetic and potential energies together with non-conservative forces provide a low-order dynamical inductive bias without requiring physiological labels or identifying latent coordinates with specific hemodynamic variables. Evaluations with frozen encoders highlight benefits for blood-pressure estimation and heart-rate estimation on WildPPG, a challenging dataset with substantial motion and environmental variation. The learned residual also supports PPG–ECG modality discrimination, providing a complementary diagnostic of compatibility with the learned dynamics. These findings suggest that physics-informed latent dynamics can complement contrastive learning for hemodynamic inference and challenging ambulatory PPG analysis.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
photoplethysmography, representation learning, self-supervised learning, physics-informed learning

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_3evAOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a PPG pretraining technique. They use contrastive and physics-informed representation learning. The authors aim to put a prior to improve PPG representation learning. The authors used twelve baselines on eight tasks to provide evidence for their technique.

Critical strengths

The overall motivation of the paper, integrating physics to the PPG representation, is strong.

The chosen datasets are good fit to the purpose of showing and validating the claims.

Critical weaknesses

The experiments seem completely wrong. WildPPG numbers, especially R^2, are worse than mean prediction (R^2 negative?). The techniques basically do not work. The numbers are also inconsistent with the published WildPPG paper. The baselines are not new. There are couple of new baselines [1] for PPG representation learning, which perform much better.

[1] Learning Without Augmenting: Unsupervised Time Series Representation Learning via Frame Projections. NeurIPS 2025.

The presentation is not clear. The paper is very hard to follow. The figures occupy too much space. Some important details in Appendix can also be brought to main manuscript.

I think the paper has significant problems in presentation and results. The reported negative R^2 values also make the reader question about the rigor of the work.

Questions for authors

How can the R^2 reach -0.7? A simple linear probe that estimates training data mean should be 0 no? The results are also completely different from the originally published work. How do the authors explain this difference?

Reviewer_RCMTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PhysPPG combines contrastive self-supervised learning with a learned dynamical regularizer for photoplethysmography (PPG). A temporal encoder maps each signal into a four-dimensional latent trajectory. Learned mass, potential-energy, and non-conservative-force networks define a generalized Euler–Lagrange residual, whose squared magnitude supplements the contrastive objective. The authors evaluate frozen representations on eight tasks across five datasets and use the residual to distinguish PPG from ECG. Results show selective improvements, particularly for blood-pressure estimation, with smaller gains on WildPPG and limited benefits for PPG-DaLiA and affective classification.

Critical strengths
  1. A well-motivated representation-learning approach. Introducing temporal dynamical structure into PPG pretraining is a reasonable direction, particularly because pulse morphology and its evolution can contain information beyond periodicity. Applying the regularizer to a time-resolved trajectory gives the method a coherent architectural rationale.

  2. Broad evaluation with a useful matched control. The experiments cover blood pressure, heart rate, and affective classification, comparing against twelve methods. The same-architecture, no-physics SimCLR control is especially valuable for assessing the additional regularizer. Separating controlled SSL comparisons from external pretrained-model comparisons also helps readers interpret the evidence.

  3. Promising improvements on blood-pressure tasks. On PPG-BP, DBP MAE decreases from 9.74 to 8.88 mmHg and RMSE from 12.27 to 11.16 mmHg relative to the matched control. PhysPPG also achieves the lowest reported SBP and DBP MAE/RMSE on PulseDB. These results support further investigation of latent dynamics for physiological representation learning.

Critical weaknesses
  1. The dynamical objective admits scale degeneracy, weakening its interpretation. Both the Lagrangian and force field are learned jointly with the encoder. Scaling LL and FncF_{\mathrm{nc}} by a common factor cc scales the residual by cc and its squared penalty by c2c^2, without changing the contrastive representation. The stated mass floor of 10−410^{-4} prevents exactly zero mass but still permits very small inertial terms, alongside nearly constant potential energy and small forces. Thus, a low residual need not demonstrate meaningful dynamical consistency. The paper should report learned mass and force magnitudes, energy gradients, and individual residual terms, and test constraints or normalization that prevent this shortcut.

  2. The experiments do not isolate the value of the Lagrangian structure. The primary ablation removes the entire physics loss. It therefore establishes the effect of an auxiliary regularizer, but cannot distinguish the proposed mechanics-based structure from generic temporal smoothing or learned dynamical prediction. Important controls include first- or second-derivative penalties, a direct latent acceleration-prediction loss, constant versus learned mass, and restricted versus unconstrained forces. Without these comparisons, the specific benefit of “physics-informed” learning remains uncertain.

  3. Statistical support for the improvements is incomplete. Several gains over SimCLR are modest: PulseDB SBP MAE improves by 0.19 mmHg, and WildPPG HR MAE by 0.80 bpm. The reported confidence intervals describe each method separately; they do not quantify uncertainty in paired performance differences. Overlapping intervals do not establish a lack of significance, but the tables also do not establish reliable superiority. Paired subject/fold analyses and multiple independent pretraining seeds would strengthen the evidence. Counting improvements on 26 of 28 metrics is less informative because MAE, RMSE, correlation, and R2R^2 are correlated measures of the same predictions.

  4. Evidence for motion robustness is indirect and limited in practical magnitude. WildPPG results remain weak, with MAE of 24.39 bpm, correlation around 0.15, and negative mean R2R^2. The improvement is encouraging, but does not yet establish reliable ambulatory HR estimation. Activity- or signal-quality-stratified evaluation and controlled corruption experiments would help determine whether the regularizer specifically improves resistance to artifacts. The poorer ranking on PPG-DaLiA further limits general robustness claims.

  5. The residual experiment demonstrates modality discrimination rather than useful PPG quality assessment. AUROC of 0.986 for PPG versus ECG is striking, but ECG differs substantially in waveform morphology. This experiment does not determine whether the score detects motion-corrupted PPG, poor contact, device shifts, or physiological abnormalities. Comparisons with simpler morphology, spectral, and feature-distance scores are needed. The manuscript should also specify the data sources, subject separation, preprocessing, threshold-selection procedure, and training budget for the post-hoc control dynamics module.

  6. Some protocol details remain insufficiently specified. The text describes PulseDB evaluation as LOSO, while Table 3 reports 20 evaluation folds; their relationship should be clarified. The PPG-BP five-fold protocol should explicitly state whether all recordings from each participant remain in one fold. Exact pretraining subset sizes, handling of held-out PulseDB subjects across evaluations, latent temporal resolution, and model-specific adapters are also needed to assess reproducibility and comparison fairness. Supplementary code is mentioned, but was not available in the supplied material.

Questions for authors

Please see the weakness part.

Reviewer_kfXXOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work introduces PhysPPG for self‑supervised PPG representation learning. It builds upon contrastive learning and adds a physics‑based regularizer derived from the generalized Euler‑Lagrange residual. The encoder outputs time‑dependent latent trajectories. These trajectories are regularized with learned mass, potential‑energy, and non‑conservative‑force components. The authors run experiments across five datasets and eight downstream tasks under a frozen‑encoder setup. They observe improvements for blood‑pressure estimation and heart‑rate estimation on the challenging WildPPG dataset. However, gains do not appear across every task.

Critical strengths
  1. The proposed modeling design, pairing temporal latent trajectories with an energy-and-force structure, offers a novel way to regularize self-supervised PPG representations.
  2. Under frozen-encoder evaluation, the method reduces prediction errors on blood-pressure and heart-rate estimation tasks.
  3. The theoretical derivation for the generalized Euler–Lagrange residual is clearly presented and appears mathematically sound.
Critical weaknesses
  1. The benefit of this specific modeling choice is not sufficiently demonstrated. Learnable Lagrangian dynamics and physics-informed learning already have prior work. The paper motivates applying these ideas to PPG, but its no-physics baseline lumps smoothed differentiation, dynamical constraints, and the energy–force networks together. This makes it hard to separate gains from the Lagrangian structure versus generic temporal regularization. The work lacks such experiments.
  2. A potential scale degeneracy in the physics loss remains unaddressed. Mass, potential energy and force are all learned quantities. If mass nears its small constant lower bound, potential energy becomes nearly constant and force drops close to zero. The residual can shrink significantly without altering the latent trajectory. The non-collapse analysis in Appendix E does not directly rule out this scenario.
  3. Ranking against pretrained models lacks sufficient supporting implementation details. The paper mentions model-specific adapters and matched downstream parameter counts, but it does not report each model’s input window length, sampling rate, temporal feature resolution or adapter design. These implementation choices can change performance under the shared temporal probe setup. It is also unclear how PulseDB test-subject exclusion rules apply to the released external checkpoints.
Questions for authors
  1. Could you provide ablation experiments that separate gains from the Lagrangian structure from generic temporal regularization? Please keep the encoder, pretraining data, and tuning budget matched across conditions. These comparisons would help establish whether the observed gains stem from the proposed physics formulation.
  2. What prevents the physics loss from decreasing simply by reducing the scale of learned dynamics? Please report the learned mass distribution and typical magnitudes of inertial, potential-gradient, and force terms. Do the observed gains remain when mass scale is fixed or normalized? Nonzero latent magnitudes alone do not resolve this concern.
  3. Please explain whether input and feature configurations were aligned across all compared pretrained models. How were models with pooled outputs or different temporal resolutions connected to the shared temporal probe?
  4. Please clarify how test subjects are excluded from pretraining on PulseDB. For the released external checkpoints used for comparison, was overlap checked between their pretraining data and these PulseDB test subjects?
1 comment

Hi authors, I'm looking forward to discussing the paper. If my concerns are addressed, I will reconsider my score.

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

Related papers

Discussion (0)

Sign in to comment.