acceptodds
Withdrawn from ICLR 2027

Individual-level Public-Private Decoupling for Continuous Blood Pressure Waveform Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing methods for continuous arterial blood pressure (ABP) waveform prediction from photoplethysmography (PPG) do not adequately distinguish inter-individual differences in blood pressure. To address this limitation, we propose Individual-level Public-Private Decoupling (IPPD), a network that decomposes PPG representations into statistically decorrelated Public and Private features and expresses the ABP prediction as the sum of a public waveform and an individual residual. To better distinguish within-individual consistency from between-individual variation, we introduce dual-view contrastive learning into the Private branch. On VitalDB-360, IPPD achieves the best individual-level prediction performance. Without target-case calibration or fine-tuning, it attains a waveform RMSE of 11.65 mmHg and a waveform MAE of 9.91 mmHg across 30 held-out test cases, outperforming existing blood-pressure prediction baselines. The code will be released.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Public-Private Decoupling, Continuous Blood Pressure Estimation, Individual-specific Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Q2FjOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies continuous arterial blood pressure waveform prediction from a single photoplethysmography window on previously unseen cases. Papers falling under the category of AI in interdisciplinary healthcare, identifying scientific problems within the medical field and seeking solutions through AI. The authors employ a shared ResNet-1D backbone followed by "common" and "private" branches, fusing their waveform outputs via smooth convolution. Training adds Public-waveform supervision, a representation-difference penalty, contrastive learning between two augmentations of a window, and training-case identity classification. The evaluation uses a curated VitalDB subset with 300 training, 30 validation, and 30 test cases. The paper reports case-macro-averaged waveform and pressure-statistic errors against four reproduced baselines and a sequential component ablation.

Critical strengths

As seen on page 6, section 4.1 of the paper, they used complete case data instead of randomly split data windows to divide the samples, which is good because this approach aligns with the generalization problem discussed in the paper. Furthermore, they incorporated several methods, such as public-head supervision and private-head zero-initialization methods.

Table 1 reports improvements across all seven displayed metrics, and the percentage reductions in Section 4.3 agree with the table. Table 2 also reports monotonically decreasing waveform errors. The predictor does improve on the reported metrics. The missing controls, discussed below, make it hard to tell why.

Critical weaknesses

In my assessment, the work in its current form is not sufficiently developed to support its central claims. The missing validation concerns the proposed Public/Private separation itself, and the conflicting method descriptions leave me unable to determine exactly what was evaluated. Stronger experimental controls are also needed to establish whether the numerical gains amount to an independent empirical contribution. Addressing these issues would require substantial additional work before I could consider the paper ready for acceptance.

  1. My main concern is that the experiments do not demonstrate the separation on which the paper bases its contribution (Abstract; contributions on p2; Sections 3.3-3.6; Table 2; Conclusion). Shared/private encoders and difference penalties already appear in Domain Separation Networks, and subject-shared/private physiological representations appear in the cited MDNet. IPPD differs by using an input-dependent Private branch to correct an unseen case's waveform. I would like to see evidence of what this adaptation adds.

    The Public head is supervised against each case's complete ABP waveform, which contains individual variation. The identity classifier encourages the Private branch to retain case information, but that alone does not show that the two branches learn the proposed roles. A cross-covariance penalty constrains second-order dependence; it does not identify which branch contains reusable physiology or individual variation. The experiments do not test even the weaker claim that one branch is relatively more shared and the other more sensitive to the individual. The contrastive positives are two augmentations of one window, so their agreement does not establish stability across different windows from an individual.

    Table 2 does not resolve this: the Private branch brings both extra capacity and another loss, and the remaining rows test just one sequence of component additions. There is no fixed-capacity removal of the difference penalty, separate test of Public-head supervision, or direct branch-role diagnostic. The conclusion leaves role stability and interpretability to future work (p8), even though the paper already relies on those roles to explain its results. A capacity-matched residual control and frozen-feature comparisons of case information and within-case stability would test the existing claim. Without those tests, I cannot attribute the gain to the proposed decoupling or judge whether it is needed.

  2. I could not determine exactly which model and loss produced the results (Figure 1 and Section 3.4, p4; Equations 6-7, p5; Sections 3.7-4.2, p6). Figure 1 calls the Private encoder a Transformer; Section 3.4 says it has the same two-layer convolutional architecture as the Public encoder. These are materially different models.

    Equation 6 calls its expression a cross-covariance estimate without specifying centering or reshaping. If the intended inputs are centered BT-by-C matrices, the extra division by B makes the squared penalty 1/B² times the usual mean squared sample cross-covariance: a factor of 65,536 at the stated batch size of 256. This calculation assumes the interpretation above and applies to the written formula; I do not have the implementation. The missing details matter because they change the strength of the decoupling penalty. Equation 7 also leaves the view indexed by q_j undefined, and the contrastive temperature and pooling parameter k are unspecified.

    The paper further uses standardized PPG, normalized ABP, and final errors in mmHg without defining the transformations or their inverse. To assess the calibration-free claim, I need to know whether absolute pressure can be recovered using only statistics available at inference. These omissions conflict with the statement on p8 that the full architecture and hyperparameters are provided. The authors need to provide the executed architecture, exact loss operations and scales, and normalization protocol before the method can be reproduced faithfully.

  3. The empirical evidence is also insufficient to justify the contribution on predictive performance alone (Sections 4.1-4.4, pp6-7). The closest reported baseline, PaPaGei, is worse by 0.5564 mmHg in waveform RMSE and 0.5153 mmHg in MAE; the KS difference is 0.0023. Table 1 does not define its “±” values, and the manuscript supplies neither a run count nor paired case-level uncertainty. I therefore cannot tell how consistently the gains hold across runs or cases.

    The reproduced baselines' waveform adaptations, pretrained-weight use, and model-specific training and selection procedures are insufficiently described. The paper cites a directly relevant waveform method, INN-PAR, ICASSP 2025, but omits it from evaluation. That work also studies subject-disjoint prediction without subject calibration. The authors should explain how IPPD differs and, where feasible, compare the two under the same VitalDB protocol. Its published numbers from other datasets cannot substitute for that comparison. To judge the practical gain, I would need paired case-level results and enough detail to check how the baselines were trained. A direct waveform method evaluated under the same protocol would also help. The current table is not enough to support the broader claim of better individual-level modeling.

I will not change my score in this round.

Questions for authors
  1. Does the gain survive removal of the difference penalty with capacity, augmentation, and other losses fixed? Can branch probes show the proposed relative shared/individual roles across distinct windows? Positive results would directly address the main contribution objection and could raise Contribution to 2 or higher.

  2. Which Private encoder generated the tables, and what tensor operations implement Equation 6? The executed specification and loss magnitudes would resolve the contradiction; a materially different regularizer would require corresponding corrections to the claims and results.

  3. What are the PPG/ABP normalization and inverse transformations, and how were the 360 cases and splits selected? An inference-valid protocol without target ABP statistics would substantiate the calibration-free interpretation. I have not established that leakage occurred.

  4. What do the uncertainty values represent, how were baselines adapted, and do paired case-level gains persist against a matched direct waveform method? Consistent gains under documented, comparable training would improve my assessment of the empirical contribution, while leaving the mechanism question to be answered separately.

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

This paper proposes a method of reconstructing the arterial blood pressure (ABP) from photoplethysmography (PPG) in a way that accounts for inter-subject variability. The proposed IPPD model comprises two branches: a public branch dedicated to waveform reconstruction, and a private branch that adds subject-specific corrections based on the subject's identity. At test time, the model only requires unlabeled PPG segments without any information about subject identity.

Critical strengths
  • I appreciate how clearly and concisely the paper is written. This made it very easy to digest the key contributions with minimal excess
  • The proposed architecture is intuitive and well-justified, with the public branch anchoring a shared baseline and the private branch handling individual variation.
Critical weaknesses

Although I find the proposed architecture interesting, my key concern is that the experimental design does not do enough to support the authors' claims or to show that the model can track meaningful changes in blood pressure.

SUPPORTING GENERALIZABILITY CLAIMS:

  • One of the key claims made by this paper is that IPPD is able to learn how to generate subject-specific adjustments without fine-tuning or calibration. This means it is important for the authors to show how their model generates predictions on unseen individuals. Although the experiments are properly designed to demonstrate such generalization by creating train-validation-test splits separated by subjects (assuming case == subject), the test size of only 30 subjects is extremely modest (<10% of the total dataset). The conclusions drawn from this test set are unlikely to be robust or representative, particularly if the authors themselves assert that there is high inter-individual variability in physiological signals. Therefore, I encourage the authors to run full cross-validation to strengthen their claims about generalizability.
  • Is my understanding correct that the models evaluated for the ablation study were trained from scratch only using the corresponding model components? If so, I would also like to see the performance of a model trained with all of the components but zeroing out the private branch at test time. Comparing this with the public-only baseline in the ablation study would confirm that the public branch in the full IPPD model architecture is actually learning a baseline waveform.
  • Along those lines, it would be great for the authors to provide some example waveforms visualizing the outputs, especially if they can be separated by the public and private branches. This would provide supporting evidence that the two branches are accomplishing what the authors are claiming.

EVALUATION METRICS:

  • Waveform RMSE and MAE are reasonable metrics for tracking the accuracy of waveform reconstruction. Waveform KS (Kolmogorov–Smirnov?) is something that I have not seen as often and feels less appropriate since it only compares the distribution of amplitude values, thereby losing all information about phase and timing.
  • MAE for SBP, DBP, and the other vital signs is also good to see, but I would like to see standard deviations. More importantly, a model can achieve good vital sign MAE by regressing towards the mean, i.e., assuming everyone has normotensive blood pressure. However, for clinical applications like the ones targeted by VitalDB, what clinicians care about most is the ability of a model to track hypotensive and hypertensive values; MAE obscures this information. Therefore, I would like to see the authors report additional metrics that specifically show how well the model tracks non-normotensive blood pressure values. This could entail stratifying MAE by blood pressure ranges, showing Bland-Altman plots, or calculating within-subject correlation between predicted and known values. The authors could also report classification performance for detecting clinically significant thresholds, or concordance between the direction of predicted and measured changes in blood pressure
  • The authors could save space in Tables 1 and 2 by reporting performance numbers with fewer significant figures. I doubt the calculations are meaningful past 0.01 mmHg.
Questions for authors
  • The authors use the word "case" to describe how they split up their dataset. Is it correct to assume that each case belongs to a unique person? If so, it may be clearer to describe the dataset sizes by subjects rather than cases.
  • Was there any preprocessing done to ensure that the paired ABP-PPG data was valid? Having worked with datasets like VitalDB, I suspect that there are periods with low-quality or invalid readings.
  • The private branch uses amplitude scaling, baseline shifting, and temporal-speed variation as three augmentations. These are reasonable augmentations, but were others considered? I think it would be good for the authors to briefly justify why these augmentations make sense (e.g., physiological plausibility).
  • More of a suggestion than a question: I would suggest that the authors specify the window size in Section 3.1. It's mentioned in Figure 1 and Section 3.2 that the window size is 500 samples, but the sampling rate isn't mentioned until Section 4.1. This is important because it confirms that each window covers multiple pulses, thus providing useful context during model training.
  • I would like to see a secondary experiment (in an Appendix) that shows how the number of subjects in the training set affects model performance. Is the model able to learn generalizable information with fewer subjects? Is there a point of diminishing returns in adding new subjects?
Reviewer_tKQ7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work presents IPPD, a continuous ABP waveform prediction architecture with Public and Private branches. The Public branch predicts a shared baseline waveform, while the Private branch predicts an individual-sensitive correction. Their fusion yields the final individual-level ABP waveform prediction. The authors develop a dual-view contrastive learning strategy that explicitly distinguishes withinindividual consistency from between-individual variation. The authors design a joint objective for individual-level decoupling. On VitalDB-360 dataset, IPPD achieves the best individual-level prediction performance.

Critical strengths
  1. Calibration-free ABP waveform prediction for unseen individuals addresses a practically meaningful clinical problem.
  2. Compared against the strongest baseline PaPaGei, IPPD reduces waveform RMSE and MAE by 4.56% and 4.94% respectively.
  3. The step‑wise ablations show prediction error keeps decreasing as we add the private branch and its auxiliary training objectives.
Critical weaknesses
  1. The motivation for the proposed assembled architecture remains incremental. Many core ideas including public‑private decoupling and dual‑view contrastive learning have appeared in prior literature. While the paper combines these existing ideas for the PPG‑to‑ABP waveform prediction task, it does not clearly explain what distinguishes its design from prior architectures.
  2. The dual‑view contrastive objective cannot guarantee true within‑individual consistency. Positive pairs are generated by augmenting one single PPG window, rather than pairing distinct real‑world windows from the same subject. This mainly teaches augmentation robustness. Even with case‑identity classification loss, it remains ambiguous whether gains stem from learning subject‑specific signals or from simple augmentation‑based regularization.
  3. The ablation confounds decoupling effects with increased model capacity. Table 2 adds both the private branch network and decorrelation loss at the same time, making it impossible to tell how much gain comes from feature decoupling versus larger model capacity. There is also no controlled run that keeps identity classification while removing the dual‑view contrastive term. Observed step‑wise improvements only validate the full combined pipeline and cannot tease apart how much each individual mechanism contributes.
  4. Key PPG‑to‑ABP waveform baselines are missing from main results. LSCT and INN‑PAR are cited in related work but do not appear in Table 1. These task‑specific methods would serve as more direct benchmarks than generic generative models. Their absence weakens empirical evidence for how competitive IPPD really is.
Questions for authors
  1. The ablation in Table 2 mixes architectural changes and multiple auxiliary losses together, making it hard to quantify the benefit of each proposed mechanism. Could you provide additional ablation studies that disentangle the contributions from feature decorrelation, dual‑view contrastive learning, and the base two‑branch network architecture?
  2. The dual‑view contrastive loss creates positive pairs from augmented views of a single window. This mostly improves robustness to augmentations. Could you provide evidence to separate whether performance gains come from learning subject‑specific properties, or simply result from augmentation‑based regularisation?
  3. Could you include results for LSCT and INN‑PAR under the same case‑wise data split and preprocessing?
  4. Section 3.7 calculates training loss on normalized ABP waveforms. Could you provide the normalization and inverse transformation formulas, and describe how you estimate the corresponding parameters? Also confirm that inference does not use ground‑truth ABP statistics.
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.