acceptodds
Under review as a conference paper at ICLR 2027

A Physiology-Aware Hierarchical Foundation Model for Long-Context Waveform Representation Learning

14 views

Abstract

Physiological waveforms contain information at multiple timescales, from morphology within a cardiac cycle to changes across beats, minutes, and hours. Most existing waveform foundation models, however, divide signals into fixed-length patches and encode short windows independently. Consequently, their representation units may cut across meaningful cardiac structures, and their encoders cannot carry information from one window to the next. We introduce Physio-HNet, a causal hierarchical foundation model that addresses both limitations through physiology-aware dynamic chunking and continuous long-context modeling. Waveform landmarks and cardiac-cycle structure guide variable-length sub-beat and beat-level chunking, replacing fixed-length patches with physiologically informed units. Hierarchical compression and recurrent state preserve information across successive windows, allowing context to accumulate beyond isolated short segments. Following self-supervised waveform pretraining, routinely measured vital signs, including heart rate, respiratory rate, oxygen saturation, and blood pressure, supervise higher-level representations at successive measurement times along each recording. We evaluate Physio-HNet through frozen probes on two external clinical cohorts, context-length experiments, and controlled ablations. With 1.2 million parameters, the self-supervised encoder achieves performance comparable to existing waveform foundation models. Clinically anchored continued pretraining further improves blood-pressure estimation, yielding the highest and the lowest MAE in eleven of twelve blood-pressure comparisons across two cohorts and two probing protocols, including all six operating-room comparisons. Increasing available history to two hours improves estimation for most targets. Physiology-aware dynamic chunking improves several clinical estimates over fixed-length patching, while the segmentation achieving the best waveform reconstruction does not yield the best downstream representation. These findings support combining physiologically informed segmentation, continuous temporal context, and vital-sign supervision to learn transferable waveform representations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
physiological waveform,foundation models, state space models, dynamic chunking, representation learning, healthcare

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_rpMvOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors present Physio-HNet, a model that can deal with multiple time-scales for PPG. They evaluate their model on several datasets and observe strong performances when compared to baselines.

Critical strengths
  • Adapts a H-Net to PPG data.
  • comprehensive experiments
  • lots of details in the appendix.
Critical weaknesses
  • I think framing the model as "Physiology-Aware" or more for "physiological waveforms" is a bit misleading. I believe the author only tested their method on univariate PPG data, which is very specific type of data. Specifying that this model is only good for PPG data would strengthen the paper.
  • There are many instances in the paper which I think are explained in an unnecessarily complicated manner. For example in section 5.2, "Each trial starts cold, so the results do not measure uninterrupted state carry across processing windows or separate retained history from recurrent-state settling." I do not know what recurrent-state settling means, and no where in the paper mentions this term. I have to constantly make assumptions about what the author is trying to say due to inconsistencies in or overcomplicated terminology across the paper.
  • I also think this paper is not well-scoped and just has too many random experiments that are not really explained well. This is especially so in the appendix. Although i appreciate a lot of details, I think that it should be at least coherent and relevant to the paper's main topics.
  • No code. I think this is a big weakness as reproducibility is very important in architectural advancements like these. I would have loved to looked at the code to see how the Physio-Hnet is implemented and how different it is from the original Hnet.

Overall, I think this is a fancy neural network architecture with some nice benefits, however, due to the way it was communicated in the paper, it was not clear at all and I feel readers will have to spend lots of time to understand the main points of the work. Additionally, the motivation for all of the different components is not well conveyed as well.

Questions for authors
  • For clarification, is the PPG input is purely a 1d time series (not multi-variate)? I believe this is not mentioned in the paper.
  • What does lookahead mean in Figures 5 and 7?
  • line 1423 seems to have an indenting problem
  • There are many terms in the figures which are not properly explained in the paper. For example, what is a Temporary clinical head, Readout-level encoder, etc.
5 comments
AuthorsOpenReview ↗

Response to Reviewer rpMv (1/3)

Summary

We thank the reviewer for the careful review and the concrete examples. Following these comments, the revision improves the presentation. The method, data and existing results are unchanged. New results are marked as new, and new or revised text is shown in green.

  1. Scope. The paper now claims PPG representation learning only. It states its one claim, extending the context of PPG representation learning from seconds to hours, in the abstract, the introduction and the conclusion. See W1 and W3.
  2. Terminology. Terms are standardized across the main text and the appendices, and App A now lists them. See W2.
  3. Motivation. The introduction now states the problem behind each component. We also list the differences from H-Net with the purpose of each. See W4 and W5.
  4. Appendix. The appendix is reorganized to follow the paper, and an overview links each appendix to the contributions and the main-text sections it supports. See W3.
  5. Code. An anonymized repository is provided to the reviewers and the AC (https://anonymous.4open.science/r/Physio-HNet-Opensource-328F/). See W4.
  6. Figures and details. Figure labels follow the terms of the text, §3.1 states that the input is a single PPG channel, and the indentation is fixed. See Q1 to Q4.

In sum, the revision keeps the content and makes the paper easier to read, with one claim, one term per concept and code to check the implementation.

W1 Framing as "physiology-aware" / "physiological waveforms"

"Physiology-aware" and "physiological waveforms" are misleading, since only univariate PPG is tested.

We take the reviewer's suggestion and state that the model is for PPG. Every experiment uses univariate PPG. The revised abstract and introduction state the PPG scope in their first sentences. At camera-ready the title becomes "Physio-HNet: Physiologically Guided Hierarchical Representation Learning for Long-Context Photoplethysmography".

The new title says "physiologically guided" because physiological knowledge supervises the model at two scales. Waveform landmarks and the cardiac cycle guide its segmentation, and routine vital signs supervise its representation along each recording. The revised Related Work states this in §2.

W2 Overcomplicated explanations and inconsistent terminology

Explanations overcomplicated and terminology inconsistent, e.g. "recurrent-state settling" in §5.2.

"Settling" referred to the recurrent state moving away from its zero start as the input is read. The revision removes the term and rewrites the sentence.

The sentence describes the context-length experiment. Each input starts with the recurrent state set to zero, so that it contains exactly its stated context length. This is a control of the experiment. At inference there are no windows, and the state runs forward from the start of the encounter (§3.1). §5.2 now says this directly.

Before: "Each trial starts cold, so the results do not measure uninterrupted state carry across processing windows or separate retained history from recurrent-state settling."

After: "For each measurement and each context length C, the encoder receives the preceding C of recording as one input with the recurrent state set to zero at its start, so that each estimate uses exactly C of context … This zero start is a control of the experiment; elsewhere the state runs forward from the start of the encounter (§3.1)."

Beyond this sentence, we revised the terminology of the whole paper, including the appendices. App A is now a table of the terms used in the paper, and §3 points to it. The figure labels follow the same terms, as described in Q4.

AuthorsOpenReview ↗

Response to Reviewer rpMv (2/3)

W3 Scope and coherence of the experiments and appendix

Scope unclear; too many experiments, and an appendix that is not coherent.

The paper now makes one claim, and each experiment and each appendix is tied to it.

  1. Scope. The paper extends the context of PPG representation learning from seconds to hours. Cardiac-cycle hierarchies make hour-long input tractable, and vital-sign supervision teaches long-range regularities. The abstract, the introduction and the conclusion state this, and the title changes at camera-ready, as described in W1. Each main-text experiment tests one part of this claim.
    • The context-length experiment in §5.2, Figure 3 and Table 9 tests whether the representation benefits from hours of context.
    • The boundary comparison in §5.3 and Tables 3, 12 and 15 tests what cardiac-cycle supervision adds over unsupervised and fixed-stride segmentation.
    • The external comparison in §5.1 and Table 1 evaluates the frozen representation before and after continued pretraining.
    • The objective ablation in §5.4 and Table 4 explains the design choices.
  2. Appendix. The appendix is reorganized in three parts that follow the paper, namely background, method and training, and the experiments of §5.1 to §5.4. An overview links each appendix to the three contributions and the main-text sections it supports. App C argues for the design rather than testing it and is now titled as a discussion. We removed passages that were not tied to a reported result. These were the former extended discussion, the description of a decoder path that the reported experiments do not use, a repeated description of an experiment not reported here, and configuration keys copied from code.
  3. Terminology. As described under W2, the terms are now consistent across the paper, and App A lists them.

Every experiment and appendix now serves the one claim of the paper, and the overview table shows which part it supports.

W4 No code

No code; how does Physio-HNet differ from H-Net?

The training code is in an anonymized repository at https://anonymous.4open.science/r/Physio-HNet-Opensource-328F/. It contains the encoder, self-supervised pretraining, continued pretraining with vital-sign supervision, and the configurations of the four chunking variants A1, A2, B and C. A script builds each variant and runs a forward pass without data or checkpoints, and it reports the 1.2M parameters of the deployed encoder. The full code, including evaluation, will be released with the paper. For readers who want the differences from H-Net without reading the code, Physio-HNet changes five things, each for a stated purpose.

  1. Supervised boundaries. H-Net learns its boundaries end to end, with only a ratio loss on the compression rate. Physio-HNet adds landmark losses that place sub-beat boundaries at the pulse foot, the systolic peak and the dicrotic notch. A cycle loss places one beat boundary per cardiac cycle (§3.2). Each unit then describes one phase of one cardiac cycle, or one cycle, instead of mixing parts of neighbouring beats as fixed patches do. §5.3 tests this by comparing A2 with B, which uses the same router without boundary-location supervision.
  2. Each chunk is represented by its last state. H-Net represents a chunk by the vector at its start. Physio-HNet uses the causal state at the chunk's end (§3.1). This state has processed the whole unit and can retain earlier context. It is available as soon as the unit closes, so the encoder stays causal.
  3. A decoder used only in training, without an encoder skip connection. In H-Net the decoder receives a residual path from the encoder and stays part of the model for autoregressive prediction. In Physio-HNet the decoder only reconstructs the waveform during training (§3.3). Without a skip connection, each unit's representation must retain the content of its unit. The decoder runs backwards in time, so each position is rebuilt from representations that encoded it. It is discarded at inference, which leaves a 1.2M-parameter causal encoder.
  4. Recurrent state carried through the recording. In training, the Mamba-2 blocks carry their recurrent state across consecutive windows. At inference there are no windows, and the state runs forward through the whole recording. The beat sequence is about fifteen times shorter than the token sequence, so this state can span hours. This is what lets the encoder use hour-scale context.
  5. Continued pretraining with vital-sign supervision. After self-supervised pretraining, a readout level whose chunks end at vital-sign measurement times is added, and routine vital signs supervise the representation (§3.4). The readout level and supervision heads are then discarded, and the encoder is evaluated frozen.

In sum, the five changes serve one goal, a small causal encoder whose units follow the cardiac cycle and whose state carries them over hours.

AuthorsOpenReview ↗

Response to Reviewer rpMv (3/3)

W5 Overall clarity and motivation of the components

Hard to follow, and the motivation of each component is not conveyed.

The revised introduction now states the problem behind each component. The reasoning is as follows.

  1. Physiological states such as blood pressure change over minutes to hours, while released PPG foundation models encode windows of 10 to 240 s in isolation. The encoder therefore carries its recurrent state forward through the recording.
  2. Two hours of PPG span 144,000 tokens even at 50 ms per token, so the stream must be compressed. Fixed-length patches cut across cardiac cycles, so landmark and cycle losses place the boundaries of the compressed units.
  3. A unit's representation should summarize the whole unit without looking ahead, so each unit is represented by the causal state at its end. §3.1 states this reason.
  4. The compressed units must keep the content of the waveform they replace. During pretraining, a decoder with no skip connection therefore reconstructs the waveform from them alone.
  5. Objectives on the waveform alone reward modelling the waveform rather than the physiological state behind it. Continued pretraining therefore adds supervision from routine vital signs.

The three contributions follow the same reasoning, namely cardiac-cycle chunking, long-context modelling with recurrent state, and continued pretraining with vital-sign supervision. Each is tied to the experiment that tests it, as listed in W3. W4 compares the components with H-Net, and W2 describes the terminology revision.

Each component now appears in the paper as the answer to a stated problem.

Q1 Is the PPG input univariate?

Is the PPG input univariate?

Yes. The encoder takes a single PPG channel sampled at 40 Hz, a univariate time series, in both training stages and in every evaluation. §3.1 now states this explicitly.

Other physiological measurements enter training only as labels. In continued pretraining, each training window holds up to two hours of one patient's PPG (§3.4). Heart rate, respiratory rate, oxygen saturation and blood pressure supervise the representation each time they are measured within the window. This longitudinal vital-sign supervision follows the patient's physiological state as it changes over the two hours. The measurements are never inputs to the encoder. The purpose is to learn more from PPG itself. PPG reflects cardiovascular, autonomic and respiratory state only indirectly, and these labels teach the encoder which of that information to read from the waveform. If the measurements were inputs, the model could use them directly and would not need to learn this from PPG.

Q2 "lookahead" in Figures 5 and 7

What does "lookahead" mean in Figures 5 and 7?

It labels A1, a control in the boundary comparison (§5.3). App G.1 defines A1 in the paragraph "Convolutional router (A1)" and in Eq. (25). A1 places boundaries with a convolutional router whose kernel spans 15 tokens centred on the current position. Each boundary decision therefore also sees the next 7 tokens, or 350 ms, of signal. "Lookahead" referred to this. "BCE" referred to A1's per-token binary cross-entropy loss against detected landmarks, which the same paragraph states. The revised figures label it "A1", the code that App A and App G.1 define, and Figure 9 changes in the same way.

Q3 Indentation at line 1423

Indentation problem at line 1423.

The paragraph break before "Cohort and preprocessing" in App D.4 is restored in the revision.

Q4 Terms in the figures

Terms in the figures not explained, e.g. "Temporary clinical head", "Readout-level encoder".

We checked every label in Figures 1 to 12 against the text. The revised figures use the terms that the text defines and App A lists. For the two examples:

  1. The "readout-level encoder" is now the "readout encoder" (§3.4). It is added in continued pretraining. It processes the beat embeddings and divides them into chunks that end at the times the vital signs are measured. Each chunk is represented by the last beat-level state at or before its measurement time, so no later waveform is used.
  2. The "temporary clinical head" is now the "supervision heads". They estimate the vital signs at each measurement time during continued pretraining. They and the readout encoder are discarded afterwards, so only the encoder is evaluated.
Reviewer_q1yzOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Physio-HNet, a 1.2M-parameter causal hierarchical encoder for PPG built on Mamba-2 that adapts H-Net dynamic chunking to the pulse wave. Sub-beat boundaries are supervised with automatically detected landmarks (pulse foot, systolic peak, dicrotic notch), beat boundaries with a one-per-cycle objective, and a recurrent state is carried across processing windows to accumulate context. The encoder is pretrained on about 24,800 h of 40 Hz ICU PPG from one private center with waveform reconstruction ("recall"), correlation and SIGReg losses, and then "clinically anchored" by supervised continued training on MIMIC-III with heart rate, respiratory rate, SpO2 and arterial pressures as targets. Validation uses frozen static and recurrent probes on two external cohorts (MC-MED and MOVER) against four released PPG encoders, a context sweep up to two hours, a comparison of four boundary configurations, and an objective ablation on four small benchmarks. The headline result is the best R² and MAE in eleven of twelve blood-pressure comparisons after anchoring.

Critical strengths
  1. The main strength, to me, is the evaluation discipline. The two external cohorts were used in neither training stage, the authors check that no patient crosses splits (App L), verify causality by overwriting everything after the query time and getting bit-identical predictions (App. k), say openly that MC-MED is in-domain for AnyPPG, grey out SpO2 and operating-room RR with a sensible rationale, and report where their own configuration loses (A1 beats A2 in five of twelve pressure settings, App e1).
  2. A 1.2M-parameter causal encoder that streams with a carried state and is already competitive before anchoring with released PPG encoders of up to 28.2M parameters is a nice practical property for monitoring, where a model has to run continuously on modest hardware. This sounds useful to practitioners.
Critical weaknesses

There are two weaknesses that, to me, drive the recommendation, specially number 1.

  1. Novelty for a conference like ICLR and scope are very limited (i) In my opinion the methodological novelty is limited for ICLR. Each component is basically off the shelf: the dynamic chunking and the pairwise cosine router come from H-Net, the backbone is Mamba-2, the collapse regularizer is SIGReg from LeJEPA, the landmarks come from a classical derivative-based detector, carrying a recurrent state across windows with truncated backpropagation is standard practice for recurrent and state-space models, and "clinically anchored continued pretraining" is supervised multi-task fine-tuning on MIMIC-III with an SSL regularizer. What is new is supervising the router with landmark, the chunk-end convention and the causal dechunking. Even the motivation is not new at ICLR: heartlang already argued that fixed windows ignore form and rhythm, and tokenized ECG into heartbeat words and rhythm "sentences"; other works aligns ECG with text at token, beat and rhythm level. Then, to this reviewer, the contribution is engineering of known pieces for one signal. While this is valuable, I feel it fits better a venue like MLHC, ML4H, or a journal such as IEEE-JBHI. (ii) The title claims a physiology-aware foundation model for physiological waveforms, but every experiment is PPG pretrained on data from one center and one vendor, and by the authors' own description the design is for quasi-periodic waveforms with automatically detectable landmarks; other signals "would need another source of boundaries" (App a). This is definitely not generalizable as stated. If physiology is claimed, the model needs to be tested across many signals of different nature: ECG, arterial pressure and respiration, and also non-quasi-periodic signals such as EEG, EDA or EMG, where there is no beat to chunk on. Otherwise the headline should be reframed, for example to a pulse-wave or cardiac-cycle-aware PPG encoder. Appendix C argues at length about language, vision, proteins, MEG/EEG and ECG inverse problems, but none of that is tested, as the authors themselves say. (iii) Even within PPG, the paper's own analysis does not show that the deployed chunking is physiology-aware where it matters. In Table 8, the sub-beat boundaries of A2 are only +.085 F1 above a count-matched uniform grid before the clinical stage and +.019 after it, so the deployed sub-beat segmentation is close to a regular grid at the right rate. A1 has the best landmark alignment at both levels but the lowest four-task mean of the routed arms, so agreement with physiology does not track downstream utility. The paper can't both put physiology-aware chunking in the title and report that the deployed sub-beat boundaries are almost a rate-matched grid. This needs to be discussed openly and the claim narrowed, and the title… And fro publishing in ICLR, this would need to be extended. (iv) The hierarchy only spans sub-beat to beat, about one second. Everything from minutes to hours is the recurrent state of the beat backbone, which any streaming SSM or RNN provides, and the level above the beat is set by charting times, not by physiology. There is no non-hierarchical streaming baseline of the same size with the same state carry, so the value of the hierarchy is never isolated. "Continuous long-context modeling" is listed as a contribution, but App k says every trial starts cold and does not exercise state carried across windows, and App. a lists the control that would separate carried state from amount of signal as not reported. (v) The downstream tasks are concurrent estimation of six vital signs that the bedside monitor already measures directly, plus four small benchmarks. For a foundation model motivated by monitoring "where the value of the signal accumulates over hours" (App. b), I would expect tasks where long causal context actually matters, e.g., forecasting intraoperative hypotension, deterioration or AF burden, and tasks beyond vital signs. Also, the anchored encoder is never evaluated outside vital signs, although its effective dimensionality halves during anchoring; the only check reads AF at .53 to .58 AUROC before and after, which the authors themselves say is not a sensitive test. So we don't know whether anchoring narrows the representation, which matters a lot for a foundation-model claim.

  2. The empirical positioning is not broad and there are minor gains. (i) The related work and the comparison are limited and need to be considerably expanded. There is a sizable literature on Mamba and state-space models for physiological signals that is neither cited nor compared for all ppg and ecg. On the baseline side, only four released PPG encoders are run. At least a general time-series foundation model, a handcrafted-feature baseline and some mamba models are needed to know what the hierarchy and the physiology supervision add. (ii) The eleven-of-twelve blood-pressure result is for the anchored encoder, which was trained with supervision on the same quantities that are then probed. The released encoders are self-supervised or ECG-guided, not trained on blood-pressure labels, so it is not surprising that the anchored model wins on BP, and in my opinion this is not a like-for-like FM comparison; before anchoring, the SSL encoder has the highest R2 in six of twelve. It seems fair to present anchoring as a training recipe, but then the baselines should get the same continued training on MIMIC-III, and a supervised model trained from scratch with the clinical loss is needed to see how much comes from the architecture and pretraining versus from the labels. In the end, this is your selling point. (iii) The gains are often marginal and the statistics are not strong. Please report pretraining-seed variance, patient-level bootstrap intervals and paired tests instead of win counts, and a predict-the-training-mean baseline for the pressures. (iv) Each arm's objective configuration is selected on the test probes of the four auxiliary benchmarks, and the same benchmarks are then used to compare A2 with the released encoders. That comparison is optimistically biased; a held-out selection split is needed. (v) The finding that better reconstruction does not mean a better representation is presented as a result, but it is well known, and here it rests on single-seed checkpoints of arms that differ in more than one factor.

Questions for authors

My main concerns are specified above, so I copy paste them here, please answer whether you are willing to expand the scope and compare to the SoA models, this is kind of my biggest concern but also all the others. I copy paste from above, because I think rephrasing does not make sense.

  1. Novelty for a conference like ICLR and scope are very limited (i) In my opinion the methodological novelty is limited for ICLR. Each component is basically off the shelf: the dynamic chunking and the pairwise cosine router come from H-Net, the backbone is Mamba-2, the collapse regularizer is SIGReg from LeJEPA, the landmarks come from a classical derivative-based detector, carrying a recurrent state across windows with truncated backpropagation is standard practice for recurrent and state-space models, and "clinically anchored continued pretraining" is supervised multi-task fine-tuning on MIMIC-III with an SSL regularizer. What is new is supervising the router with landmark, the chunk-end convention and the causal dechunking. Even the motivation is not new at ICLR: heartlang already argued that fixed windows ignore form and rhythm, and tokenized ECG into heartbeat words and rhythm "sentences"; other works aligns ECG with text at token, beat and rhythm level. Then, to this reviewer, the contribution is engineering of known pieces for one signal. While this is valuable, I feel it fits better a venue like MLHC, ML4H, or a journal such as IEEE-JBHI. (ii) The title claims a physiology-aware foundation model for physiological waveforms, but every experiment is PPG pretrained on data from one center and one vendor, and by the authors' own description the design is for quasi-periodic waveforms with automatically detectable landmarks; other signals "would need another source of boundaries" (App a). This is definitely not generalizable as stated. If physiology is claimed, the model needs to be tested across many signals of different nature: ECG, arterial pressure and respiration, and also non-quasi-periodic signals such as EEG, EDA or EMG, where there is no beat to chunk on. Otherwise the headline should be reframed, for example to a pulse-wave or cardiac-cycle-aware PPG encoder. Appendix C argues at length about language, vision, proteins, MEG/EEG and ECG inverse problems, but none of that is tested, as the authors themselves say. (iii) Even within PPG, the paper's own analysis does not show that the deployed chunking is physiology-aware where it matters. In Table 8, the sub-beat boundaries of A2 are only +.085 F1 above a count-matched uniform grid before the clinical stage and +.019 after it, so the deployed sub-beat segmentation is close to a regular grid at the right rate. A1 has the best landmark alignment at both levels but the lowest four-task mean of the routed arms, so agreement with physiology does not track downstream utility. The paper can't both put physiology-aware chunking in the title and report that the deployed sub-beat boundaries are almost a rate-matched grid. This needs to be discussed openly and the claim narrowed, and the title… And fro publishing in ICLR, this would need to be extended. (iv) The hierarchy only spans sub-beat to beat, about one second. Everything from minutes to hours is the recurrent state of the beat backbone, which any streaming SSM or RNN provides, and the level above the beat is set by charting times, not by physiology. There is no non-hierarchical streaming baseline of the same size with the same state carry, so the value of the hierarchy is never isolated. "Continuous long-context modeling" is listed as a contribution, but App k says every trial starts cold and does not exercise state carried across windows, and App. a lists the control that would separate carried state from amount of signal as not reported. (v) The downstream tasks are concurrent estimation of six vital signs that the bedside monitor already measures directly, plus four small benchmarks. For a foundation model motivated by monitoring "where the value of the signal accumulates over hours" (App. b), I would expect tasks where long causal context actually matters, e.g., forecasting intraoperative hypotension, deterioration or AF burden, and tasks beyond vital signs. Also, the anchored encoder is never evaluated outside vital signs, although its effective dimensionality halves during anchoring; the only check reads AF at .53 to .58 AUROC before and after, which the authors themselves say is not a sensitive test. So we don't know whether anchoring narrows the representation, which matters a lot for a foundation-model claim.

  2. The empirical positioning is not broad and there are minor gains. (i) The related work and the comparison are limited and need to be considerably expanded. There is a sizable literature on Mamba and state-space models for physiological signals that is neither cited nor compared for all ppg and ecg. On the baseline side, only four released PPG encoders are run. At least a general time-series foundation model, a handcrafted-feature baseline and some mamba models are needed to know what the hierarchy and the physiology supervision add. (ii) The eleven-of-twelve blood-pressure result is for the anchored encoder, which was trained with supervision on the same quantities that are then probed. The released encoders are self-supervised or ECG-guided, not trained on blood-pressure labels, so it is not surprising that the anchored model wins on BP, and in my opinion this is not a like-for-like FM comparison; before anchoring, the SSL encoder has the highest R2 in six of twelve. It seems fair to present anchoring as a training recipe, but then the baselines should get the same continued training on MIMIC-III, and a supervised model trained from scratch with the clinical loss is needed to see how much comes from the architecture and pretraining versus from the labels. In the end, this is your selling point. (iii) The gains are often marginal and the statistics are not strong. Please report pretraining-seed variance, patient-level bootstrap intervals and paired tests instead of win counts, and a predict-the-training-mean baseline for the pressures. (iv) Each arm's objective configuration is selected on the test probes of the four auxiliary benchmarks, and the same benchmarks are then used to compare A2 with the released encoders. That comparison is optimistically biased; a held-out selection split is needed. (v) The finding that better reconstruction does not mean a better representation is presented as a result, but it is well known, and here it rests on single-seed checkpoints of arms that differ in more than one factor.

12 comments
AuthorsOpenReview ↗

Response to Reviewer q1yz (1/5)

Summary

We thank the reviewer for the constructive assessment and for recognizing our careful evaluation and the practical value of our efficient model for continuous PPG analysis. To address concerns about novelty, scope, and empirical support, we have expanded the comparisons, added analyses and a clinically relevant forecasting task, and applied continued pretraining to AnyPPG. We also strengthen the statistical analysis and narrow the manuscript's scope to PPG, as the reviewer suggests.

The reviewer asks whether we are willing to expand the scope and compare with state-of-the-art models. We have done both within PPG.

  1. Comparisons. We add Chronos and MOMENT, two general time-series foundation models, and a pyPPG handcrafted-feature baseline. We also apply continued pretraining to AnyPPG, the strongest released encoder, and train the same architecture from scratch with the clinical objective alone.
  2. Tasks. We add intraoperative hypotension forecasting, a task where long causal context matters. We also evaluate the encoder after continued pretraining on tasks beyond vital signs.
  3. Signals. Every experiment uses PPG, so the paper now claims PPG representation learning only, the reviewer's second option in (ii). The title changes at camera-ready.

Together, these additions provide stronger evidence for boundary accuracy, long-context processing, and transfer to clinical tasks. They also clarify which contributions are directly supported by controlled experiments and which remain limitations.

The revision calls clinical anchoring "continued pretraining with vital-sign supervision", and "+ vitals" marks the encoder after it. The arms A1, A2, B and C are now called chunking variants. Our response follows the reviewer's two main points, and (i) to (v) follow the reviewer's numbering. Part 1 states what is new and then answers the questions about the evidence. Part 2 lists the added experiments and what each one shows.

AuthorsOpenReview ↗

Response to Reviewer q1yz (2/5)

Part 1. Novelty and scope

What is new

Although Physio-HNet incorporates established components, it introduces a new 1.2M-parameter causal PPG encoder that learns physiologically guided units and accumulates context over hours of recording. This design extends PPG representation learning from independent short windows to continuous, long-context encoding. Released PPG foundation models encode windows of 10 to 240 s independently.

Why hours matter. Physiological states such as blood pressure change over minutes to hours. The new hypotension forecasting task tests whether this longer context helps (new, see (v)). Physio-HNet carries its recurrent state through at least one hour of recording, while the other encoders read one window of 10 to 240 s at a time. Physio-HNet has the highest AUROC at 5, 10 and 15 minutes, and after continued pretraining also the highest AUPRC. The encoder that carries hours of context thus forecasts this clinical event best.

Reaching hours raises four problems, and each one shapes a part of the design.

  1. Compression. Two hours of PPG are 144,000 tokens, and a flat streaming model updates every layer at this rate. Its fixed-size state cannot keep everything it has seen (Jelassi et al., 2024), and its effective receptive field is limited by the training length (Ben-Kish et al., 2025). Physio-HNet compresses the waveform into one embedding per cardiac cycle, so its beat backbone makes about 9,400 updates in two hours.
  2. Where to cut. Fixed-length patches cut across cardiac cycles. HeartLang cuts beats with a QRS rule before training, and MELP pools each 10 s ECG into tokens without explicit boundaries. Both are now cited, and both stay within one 10 s window. In Physio-HNet the router learns the boundaries from the encoder states, and landmarks and cardiac cycles supervise it only in training. The boundary losses update the same states that become the unit representations. Learning where to cut therefore also shapes what each unit represents. The chunking variants test this. A2 and B differ only in boundary supervision. After continued pretraining A2 transfers better (Tables 3 and 12), and its four-task mean is also higher (Table 4). A2 and C share the encoder stacks, state carry and unit rate but form units differently. A2 is above C on every blood-pressure target in all four external settings (Table 12). The new K0-only and K1-only variants show that the two levels of supervision work together (Table 12, and our response to Reviewer 7t9Q, Q1).
  3. Content and causality. Each unit is represented by the causal state at its end. This state has processed the whole unit and is available as soon as the unit closes. The decoder is used only in training and has no skip connection from the encoder, so each embedding must carry the content of its own segment. The recurrent state is carried across training windows, so long-range learning starts in pretraining, before any labels are used. The decoder without a skip connection and the carried state are two changes to H-Net beyond the three the reviewer lists. Our response to Reviewer rpMv, W4, lists all five differences from H-Net with their purposes.
  4. Physiology behind the waveform. Objectives on the waveform alone reward modelling the waveform rather than the physiological state behind it. Continued pretraining uses the same procedure as supervised multi-task fine-tuning, but with a different role. It supervises the representation with routine vital signs at their measurement times along two-hour recordings. The layers and heads added for this stage are then discarded, and the frozen encoder is probed on two external cohorts.

Signal-specific representation learning has appeared at ICLR, for example PaPaGei and HeartLang at ICLR 2025. How to form units and carry context in physiological streams is a representation-learning question.

The revised manuscript makes these contributions and their relationship to existing methods explicit.

AuthorsOpenReview ↗

Response to Reviewer q1yz (3/5)

Part 1. Novelty and scope (continued)

Questions about the evidence

(ii) Scope and title. We agree that the demonstrated scope is PPG. The revised abstract, introduction and limitations state this scope. We also propose the title:

“Physio-HNet: Physiologically Guided Hierarchical Representation Learning for Long-Context Photoplethysmography.”

Extension to ECG, arterial pressure, and other signals remains future work. Signals without cardiac-cycle structure would require a different source of representation units. Appendix C is now presented as a discussion motivating the design, rather than evidence of generalization across modalities.

(iii) Sub-beat boundaries and a regular grid. A2's sub-beat boundaries are not close to a regular grid. The small difference in Table 15 comes from a loose matching criterion. Table 15 counts a boundary as matching a landmark if it lies within 100 ms. Sub-beat landmarks, namely the foot, the systolic peak and the dicrotic notch, are only about 250 ms apart, so almost any boundary lies within 100 ms of one. We re-scored the same windows with stricter criteria (new). The table gives the F1 of A2 and of a regular grid with the same number of boundaries. SSL denotes A2 before continued pretraining.

Counted as a match if withinA2, SSLGridA2, + vitalsGrid
100 ms (Table 15).846.751.833.776
50 ms.839.519.827.557
the same 50 ms token.824.180.814.193
  • On the landmark's own token, A2 still scores .82, while the grid falls to .18, close to random placement at .17. §5.3 and Table 15 now report this F1.
  • At 100 ms the difference shrinks after continued pretraining, because A2 then places more boundaries, 277 rather than 231 per minute. The matched grid becomes denser and scores higher, while A2's own F1 barely changes.
  • The probes read the beat embeddings (App E). Beats lie about 770 ms apart. At this level A2's F1 is .58 above the matched grid before continued pretraining and .63 after, against about zero for C.
  • A1 places boundaries more accurately than A2 but has a lower four-task mean in Table 4, .784 against .812. A1's router is a separate convolutional network that can locate landmarks by itself. A2's router compares two adjacent encoder states, so the encoder states themselves must change at a landmark. This suggests that boundary supervision helps through the encoder states it shapes.

Together, A2's sub-beat boundaries fall on the landmarks, and its beat boundaries follow the cardiac cycle.

(iv) Hierarchy and long context. We agree that context over minutes to hours is carried by the recurrent state of the beat backbone. The hierarchy decides what that state carries, one embedding per cardiac cycle instead of one token every 50 ms, so its value lies in the units the state is built from.

  1. A flat streaming baseline. A flat model cannot be pretrained with our objective. Our pretraining reconstructs the waveform from the compressed embeddings. Without compression, the output at each token already contains that token, so reconstruction reduces to copying the input. A flat baseline would therefore change the objective as well as the hierarchy. The variants instead change only the units. C keeps the hierarchy, the state carry and the unit rate of A2 but cuts at a fixed stride, and B keeps A2's router without boundary supervision. A2 is above C on every blood-pressure target in all four external settings, and at or above B after continued pretraining (Table 12). With the same state carry, what the state carries matters.
  2. The level above the beat. It follows charting times and is not part of the hierarchy we claim. It is added only during continued pretraining, to place vital-sign supervision at measurement times (§3.4). The probes read the beat embeddings and do not use the readout level (App E).
  3. Trials that start from a zero state. At inference there are no windows. The encoder is causal and recurrent, so its state runs forward through the whole recording. A two-hour input in the context-length experiment (§5.2, App F.1) therefore uses the state over two hours. The zero state marks where the evaluated recording begins, so each estimate uses a known amount of context.
Reviewer_7t9QOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Physio-HNet is a PPG foundation model designed to capture both pulse shape and changes over longer recordings. Waveform landmarks and cardiac-cycle structure guide its sub-beat and beat-level chunking during training, while causal Mamba carries information across processing windows. After self-supervised waveform pretraining, the model continues training on MIMIC-III using routinely measured vital signs. The authors evaluate the frozen encoder on MC-MED and MOVER and examine context length, boundary supervision, and training objectives. The clinically anchored model achieves the highest mean R² and lowest mean MAE in 11 of 12 blood-pressure comparisons across the two cohorts and two probing protocols.

Critical strengths
  1. Evaluation uses two external clinical cohorts not used to pretrain Physio-HNet and two probing protocols. The clinically anchored model achieves the highest mean R² and lowest mean MAE in 11 of 12 blood-pressure comparisons. Context-length and objective comparisons broaden the evaluation, and the A2–B comparison with the same backbone supports boundary-location supervision.
  2. The two chunking levels reflect local pulse morphology and cardiac-cycle structure. Hierarchical compression shortens the sequence, while causal Mamba carries history across windows. This is a suitable design for modeling continuous PPG across multiple beats.
Critical weaknesses
  1. The paper does not report how other encoders perform after continued pretraining on the same MIMIC-III data with the same vital-sign supervision. Strong baselines might also benefit from similar adaptation, limiting attribution of the final cross-model advantage to the encoder design. The self-supervised results and A2–B comparison still support the design itself.
  2. Table 3 compares pretraining-objective configurations, using their four-task mean to select each arm's configuration for clinical continued pretraining. The selected configuration's auxiliary-task scores may therefore reflect selection effects; no separate test results independent of selection are reported. This is a secondary concern about those scores, since Table 1 provides independent external clinical evidence.
Questions for authors
  1. How do landmark supervision at K0 and cardiac-cycle supervision at K1 contribute separately? The A2–B comparison changes both levels together, so their individual effects remain unclear.
  2. During clinically anchored continued pretraining, how does retaining the self-supervised objectives affect transfer to external cohorts compared with vital-sign supervision alone? Are corresponding control results available?
  3. Under the same hardware conditions, what are the runtime and memory requirements for continuously processing one minute of PPG, compared with fixed-stride chunking using the same backbone?
7 comments
AuthorsOpenReview ↗

Response to Reviewer 7t9Q (1/5)

Summary

We thank the reviewer for the careful reading and for noting that the external cohorts, the A2–B comparison and the design for continuous PPG support the work. The revision makes the main claim explicit, extending the context of PPG representation learning from seconds to hours. It also narrows the scope to PPG. The arms A1, A2, B and C are now called chunking variants. In response to this review we add the following.

  1. AnyPPG, the strongest released encoder, after continued pretraining on MIMIC-III. See W1.
  2. Continued pretraining repeated without the self-supervised terms. It separates the effect of the self-supervised objectives from that of vital-sign supervision. See Q2.
  3. K0-only and K1-only variants, which complete a 2×2 design with A2 and B. See Q1.
  4. Runtime and working memory for one minute and two hours of PPG, for A2 and C. See Q3.
  5. The encoder after continued pretraining evaluated on the four benchmarks of Table 4, and a statement of what Table 4 tests in §4 and §5.4. See W2.

In sum, each question is answered with a new measurement, and the external evidence of Table 1 does not depend on any of the choices discussed below.

AuthorsOpenReview ↗

Response to Reviewer 7t9Q (2/5)

W1 Baselines without the same MIMIC-III training

Other encoders not given the same continued pretraining on MIMIC-III, so the cross-model advantage cannot be attributed to the encoder design alone.

We now apply continued pretraining on MIMIC-III to AnyPPG, the strongest released encoder (new). After this training, Physio-HNet is still higher on blood pressure in the operating room under the static probe and in the emergency department under the recurrent probe. The patient-level bootstrap intervals of the differences exclude zero for diastolic pressure in the first and systolic pressure in the second, and AnyPPG is not significantly higher in any setting (Table 8, App E, and our response to Reviewer q1yz, Part 2 (iii)). As the reviewer notes, the self-supervised results and the A2–B comparison also support the design itself. §5.1 now reports this comparison, and the abstract states its result.

The table compares AnyPPG and Physio-HNet before and after continued pretraining. It gives R² on the two external cohorts of Table 1, as the mean ± standard deviation over five probe seeds. MC-MED is from the emergency department and MOVER from the operating room. "+ vitals" marks the encoder after continued pretraining. As in Table 1, SpO2 and operating-room RR are not used to compare encoders (§5.1).

SettingEncoderHRSpO2RRNBP-SNBP-DNBP-M
Static, MC-MEDAnyPPG.891±.010.059±.017.112±.032.254±.015.184±.026.227±.018
AnyPPG + vitals (new).884±.017.057±.018.126±.020.259±.019.209±.020.229±.015
Physio-HNet SSL.900±.006.006±.052.066±.081.241±.027.173±.017.179±.014
Physio-HNet + vitals.898±.007.022±.042.204±.053.269±.022.195±.023.219±.024
Static, MOVERAnyPPG.644±.038−.073±.025.020±.009.097±.047.108±.070.117±.059
AnyPPG + vitals (new).661±.006−.086±.032.022±.016.118±.026.119±.043.134±.036
Physio-HNet SSL.687±.009−.059±.033.004±.037.089±.082.140±.042.123±.060
Physio-HNet + vitals.696±.008−.059±.029.045±.021.149±.012.165±.010.166±.008
Recurrent, MC-MEDAnyPPG.920±.006.056±.013.134±.040.252±.024.198±.022.212±.020
AnyPPG + vitals (new).916±.008.055±.011.127±.032.262±.009.210±.015.223±.016
Physio-HNet SSL.916±.003.061±.015.119±.030.269±.014.180±.025.203±.020
Physio-HNet + vitals.920±.004.068±.020.217±.034.303±.015.220±.006.242±.008
Recurrent, MOVERAnyPPG.680±.006−.072±.036.031±.007.127±.036.146±.052.154±.041
AnyPPG + vitals (new).681±.004−.073±.031.021±.012.154±.016.167±.020.179±.012
Physio-HNet SSL.671±.003−.055±.027.016±.023.138±.013.166±.018.162±.018
Physio-HNet + vitals.671±.004−.059±.026.059±.006.158±.012.184±.016.181±.015
  1. Continued pretraining for AnyPPG (new). It uses the same MIMIC-III data and training schedule as ours. After continued pretraining, AnyPPG's encoder is frozen and probed as in Table 1.
    • The training raises AnyPPG's blood-pressure R² in every setting.
    • Under the static probe, Physio-HNet stays ahead in the operating room, a cohort no encoder was pretrained on. It has 1.2M parameters against AnyPPG's 4.04M.
    • In the other two settings the two encoders are close. AnyPPG is slightly higher on diastolic and mean pressure in the emergency department under the static probe. This cohort supplies about 72% of AnyPPG's pretraining data.
  2. Our encoder stays close to its self-supervised form. During continued pretraining, a readout level and supervision heads are added on top of the beat backbone. They read the beat embeddings and predict the vital signs. Both are discarded afterwards, and the downstream probes read the beat embeddings directly (Figure 2c). The self-supervised objectives are retained, and one recurrent state runs through each two-hour recording. The relative weight change of the encoder is therefore 2.7 to 5.3%. The vital signs help the encoder interpret the PPG.

In sum, continued pretraining alone does not explain the cross-model advantage, because Physio-HNet remains level with or above AnyPPG when AnyPPG receives the same continued pretraining. Trained from scratch on the vital signs alone, the same architecture reaches blood-pressure R² near zero (our response to Reviewer q1yz, Part 2 (ii)). The self-supervised rows of Table 1 and the A2–B comparison show what the design itself adds.

AuthorsOpenReview ↗

Response to Reviewer 7t9Q (3/5)

W2 Selection in Table 3

Table 3's (now Table 4) four-task mean selects each arm's configuration, so the selected scores may carry selection effects (a secondary concern).

Table 4 ablates the self-supervised objectives within each variant, and every variant's objective configuration is chosen by the same rule. Table 4 supports one claim, that the encoder still works on these benchmarks, and the external evidence of Table 1 does not depend on the choice.

  1. Table 4 tests three things that Table 1 does not.
    • Generalization. Its benchmarks include wrist PPG from wearables, in WESAD and the Stanford AF cohort. Its tasks are emotion, arrhythmia, activity and identity, which are unrelated to the vital signs used in continued pretraining.
    • Short windows. Its segments of 6 to 25 s match the setting the released encoders are built for, since they encode windows of 10 to 240 s. An encoder built for hour-long context also works there. A2's four-task mean, .812, is level with that of the best released encoder, AnyPPG at .804, within the reported intervals (§5.4).
    • Objectives within each variant. The same objective acts differently on different boundary mechanisms. Removing next-beat prediction raises the four-task mean of A1, A2 and B, which place boundaries with a router. It lowers that of the fixed-stride variant C, from .785 to .680. With a router, boundaries follow the content, and reconstruction alone keeps each unit's content. Fixed-stride units do not follow the cardiac cycle and need next-beat prediction to organize the representation.
  2. Each variant keeps the objective configuration that suits its boundary mechanism, in continued pretraining and in all later evaluations. One configuration for all would suit either the variants with a router or C, so the choice serves a fair comparison between variants. The configuration must serve two-hour context while its embeddings stay meaningful on short segments. Among the causal variants, A2 leads at both scales. Its four-task mean is the highest in Table 4, .812 against .797 for B and .785 for C. On the external cohorts after continued pretraining, A2 is at or above B and C on every blood-pressure target (Table 12).
  3. Table 1 does not depend on this choice. The revision also evaluates the encoder after continued pretraining on the same four benchmarks. Its four-task mean is .804, against .812 before continued pretraining and .804 for AnyPPG. The result is added to App G.2. §4 now states the purposes of Table 4, and §5.4 states this reasoning.

In sum, each variant's objective configuration comes from the same ablation within each variant, Table 4 shows that the encoder still works on these benchmarks, and the external evidence in Table 1 does not depend on the choice.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of changes

We thank the reviewers and the AC for their time and constructive feedback. The revision states one claim, extending the context of PPG representation learning from seconds to hours, and narrows the scope to PPG. It adds the experiments and statistics the reviewers asked for. Revised text is shown in green, and new results are marked as new.

  1. Claim and scope (q1yz, rpMv). The abstract, introduction, limitations and conclusion state the claim and the PPG scope. App C is presented as a discussion of the design. At camera-ready the title becomes "Physio-HNet: Physiologically Guided Hierarchical Representation Learning for Long-Context Photoplethysmography".
  2. New experiments (q1yz, 7t9Q).
    • Intraoperative hypotension forecasting, where Physio-HNet has the highest AUROC at every horizon (§5.2, Table 2).
    • Continued pretraining for AnyPPG with the same data and schedule as ours. Physio-HNet remains level or higher, and AnyPPG is not significantly higher in any setting (§5.1, Tables 1 and 8).
    • Chronos, MOMENT and pyPPG features as further baselines (Tables 1 and 4).
    • Controls within our architecture, namely training from scratch, continued pretraining without the self-supervised terms, and the K0-only and K1-only variants (App G.2, Tables 12 and 13).
    • Stricter boundary scoring. On the landmark's own token, A2 reaches F1 .82 against .18 for a rate-matched grid (§5.3, Table 15).
    • Runtime and memory of A2 against fixed-stride chunking (App F.3, Table 10).
  3. Statistics (q1yz). Patient-level paired bootstrap intervals replace the win counts (§5.1, App E, Table 8). Two further pretraining seeds of A2 will be reported in the appendix at camera-ready.
  4. Presentation (rpMv). Terms are standardized and listed in App A. The introduction states the problem behind each component, and Related Work covers HeartLang, MELP and state-space models for physiological signals. The appendix is reorganized to follow the paper and opens with an overview that links each appendix to the three contributions. The reconstruction observation is no longer presented as a finding. The training code is in an anonymized repository at https://anonymous.4open.science/r/Physio-HNet-Opensource-328F/.

The method and data are unchanged. Each change answers a reviewer comment, and our response to each reviewer gives the details.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.