acceptodds
Under review as a conference paper at ICLR 2027

Factorizing Predictable Structure and Conditional Variation for Multimodal Cardiovascular Signal Generation

Abstract

Multimodal cardiovascular signals, such as electrocardiography (ECG), photoplethysmography (PPG), and arterial blood pressure (ABP), provide complementary views of the same physiological process. Generating a missing modality from partial observations is nevertheless underdetermined: deterministic predictors tend to average plausible targets, whereas direct generative models must learn both observation-constrained structure and remaining target variation within a single distribution. We introduce a factorized framework that separates these roles in a shared latent space. Observed modalities are encoded into a global context that captures cross-modal cardiovascular structure and a local condition that retains high-resolution temporal information. A target-specific deterministic anchor first estimates the target structure predictable from the available observations. Conditional flow matching then models the residual relative to this anchor, and the final signal is obtained by decoding the sum of the anchor and a sampled residual. This formulation yields a testable hypothesis: as additional informative modalities are observed, the deterministic anchor should explain more target structure, leaving a smaller and more concentrated residual distribution. A single model supports nine generation tasks formed by single- and dual-modality observations across ECG, PPG, and ABP. We evaluate conditional accuracy, temporal alignment, distributional fidelity, and downstream physiological utility, and analyze how observation completeness affects anchor error, residual energy, and sample dispersion. The results show that separating deterministic prediction from residual generation improves distribution-level fidelity while retaining competitive conditional reconstruction, providing a practical approach to generative modeling under partial multimodal observations.

Primary area:
generative models
Keywords:
Multimodal Cardiovascular Signals, Conditional Generation, Conditional Flow Matching, Missing-Modality Generation, Physiological Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_xaHeOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CardioPRF separates conditional cardiovascular-signal generation into a deterministic prediction and a stochastic residual. Global and local conditioning pathways produce a target-specific anchor; conditional flow matching models the remaining latent variation. Their sum is decoded into the target waveform. The experiments compare this formulation with deterministic-only and direct-flow alternatives and examine reconstruction, distributional fidelity, and external downstream utility.

Critical strengths

Separating observation-predictable structure from remaining conditional variation is a clear and useful formulation of an underdetermined generation problem. The anchor–residual construction is straightforward to understand, and the complementary global/local conditions fit the structure of cardiovascular waveforms. The experiments make the trade-off between paired reconstruction and distributional fidelity visible instead of suggesting that a single metric captures all aspects of generation. The anchor-only and direct-flow comparisons are particularly informative, and the external downstream evaluations add useful context to the main results.

The formulation and empirical findings make a positive overall case for weak acceptance.

Critical weaknesses

The evidence would be easier to interpret with a sharper distinction between two pairs of claims:

  • Better distributional fidelity versus conditional or patient-specific correctness.
  • Architectural support for nine tasks versus the settings highlighted in the main tables.

These are primarily clarification and presentation issues in my assessment, rather than reasons to reject the paper.

Questions for authors

On the evaluation metrics, please specify the number of generated samples per input, how MAE is aggregated, and whether the same sampling budget is used for Min-MAE across methods. It would also help to give the precise feature space and implementation used for the reported Fréchet distance.

On task coverage, the framework supports nine directed tasks, whereas the main comparisons emphasize a subset. Could you make clear which tasks were evaluated in each table and where results for the remaining settings can be found?

Finally, on interpreting the findings, please distinguish the evidence that residual generation improves distributional fidelity from evidence about conditional physiological correctness. In particular, explain which current analyses support the hypothesis that additional observed modalities reduce unresolved variation, and frame the external downstream results within their demonstrated scope.

2 comments
AuthorsOpenReview ↗

Response to Reviewer xaHe (Part 1/2)

We sincerely thank the reviewer for the encouraging assessment and for recognizing the motivation behind predictable–residual factorization and the complementary global/local conditioning pathways. We appreciate the suggestions concerning evaluation transparency, complete task coverage, and the interpretation of conditional generation. We have addressed these points through revised metric definitions, expanded nine-task results, and additional experiments examining cardiac-rhythm consistency and inference stability.

1. Clarification of evaluation metrics and sampling procedures

We have revised Section 4.1 and expanded Appendix D to specify the evaluation metrics and sampling protocol.

For the main waveform-generation experiments, each conditioning window produces one generated realization. Standard MAE measures the direct paired waveform error, whereas Min-MAE uses bounded temporal alignment within ±1 second, following the shift-tolerant evaluation in UniCardio. The error is computed over the valid overlapping region, and the minimum across integer temporal shifts is selected for each window. Consequently, Min-MAE measures tolerance to temporal displacement rather than selecting the best of multiple stochastic samples. The same procedure applies to deterministic and stochastic models.

KS is computed from per-window empirical amplitude distributions and averaged across windows. Raw FD is calculated directly from the means and covariance matrices of the generated and recorded 512-dimensional waveform vectors, without using a pretrained feature extractor. These definitions and aggregation procedures are provided explicitly in Appendix D (Eqs. 11–15).

To further examine sampling stability, we have added five-seed inference evaluations with fixed model checkpoints (Appendix E, Tables 10–11). Each seed produces a separate full-test-set evaluation, and we report the mean and sample standard deviation of the resulting metrics. This analysis characterizes the consistency of the reported results across inference-noise realizations.

2. Complete results for all nine directed generation tasks

We agree that the original submission did not sufficiently distinguish the framework's nine-task capability from the representative tasks emphasized in the main comparisons.

The revised manuscript now explicitly reports results for all six single-observation and three dual-observation generation tasks. Appendix E provides the unified CardioPRF model's complete results (Table 7), global/local conditioning ablations (Table 8), and comparisons among anchor-only prediction, direct flow, and anchor–residual generation (Table 9).

Furthermore, Tables 10–11 extend the anchor–residual versus direct-flow comparison to all nine tasks under five inference seeds. Anchor–residual flow achieves lower mean raw FD in every task, while maintaining competitive paired reconstruction accuracy. This extends the evidence for the proposed formulation beyond the representative ECG and ABP settings in the main text.

We have also clarified the distinction between the unified nine-task CardioPRF model, the task-specific CardioPRF-F variant, and CardioPRF-M, which uses two observed modalities with the same unified model. The main comparisons remain focused on settings with established external baselines, while the appendix documents the complete task coverage and corresponding ablations.

AuthorsOpenReview ↗

Response to Reviewer xaHe (Part 2/2)

3. Conditional physiological consistency and complementary observations

We appreciate the reviewer's suggestion to distinguish waveform-distribution fidelity from conditional physiological consistency.

To complement the raw FD comparisons, we introduced a cardiac-rhythm consistency analysis for PPG→ECG generation (Section 4.4, Table 5; Appendix F). Using the same PPG observations and paired ECG references, we compare anchor-only prediction, direct flow, and anchor–residual generation in terms of cardiac-period agreement and R-peak matching.

Anchor–residual flow achieves the lowest cardiac-period errors relative to both the observed PPG (54.93 ms) and recorded ECG (60.37 ms), together with the highest R-peak micro-F1 (0.566). These results provide additional evidence that the generated ECG preserves timing information associated with its conditioning observations, complementing the distributional improvements.

Regarding additional observed modalities, the revised Table 2 shows that adding ABP to PPG reduces ECG MAE from 0.31 to 0.29, while adding ECG to PPG reduces ABP MAE from 0.60 to 0.58. These improvements indicate that complementary observations contribute useful predictive information within the shared generation framework.

We have further expanded the downstream evidence through a frozen BP-regressor experiment (Section 4.5, Table 6; Appendix H). Following the prior-work downstream evaluation protocol, five generated ECG realizations are averaged for each PPG input and supplied to the fixed predictor in place of recorded ECG. CardioPRF achieves BP estimation errors close to those obtained with recorded ECG on external Cuffless BP, providing additional evidence of downstream utility. The original external WESAD evaluation is retained in Appendix G.

The revised manuscript now more clearly connects the complete nine-task formulation, distributional fidelity, and rhythm-level physiological evidence to the corresponding empirical claims. We thank the reviewer again for the constructive suggestions and positive assessment.

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

This paper investigates conditional generation among ECG, PPG, and ABP, representing the target latent variable as a deterministic anchor predicted from observed signals and a stochastic residual generated by conditional flow matching. Global and local conditions are used to extract cross-modal context and fine-grained temporal information, respectively, and training is conducted in phases. The model is designed to support nine directed generation tasks; the main experiments demonstrate four ECG/ABP target tasks, along with downstream evaluations on two external datasets. This paper aims to demonstrate that the division of labor between anchors and residuals can improve distribution fidelity while maintaining competitive paired prediction performance.

Critical strengths
  1. The motivation aligns with the nature of the task.The paper acknowledges the uncertainty inherent in cross-modal generation and distinguishes between distribution error and distribution fidelity. This perspective helps explain why lower point prediction errors do not necessarily imply better generation results.
  2. The design corresponds clearly to the ablation results. Phased training provides a clear optimization process for both the anchor and the residual. Table 3 directly compares anchor-only, direct flow, and anchor-residual, allowing for the validation of the core design rather than merely reporting rankings against external baselines.
  3. External evaluation provides supplementary evidence. The paper explicitly states that WESAD and Cuffless BP were not used for training generative models or selecting checkpoints. Although downstream protocols still need to be supplemented, this setup helps examine the utility of generative signals beyond the training distribution.
Critical weaknesses
  1. The relationships between key tables cannot yet be verified as described in the text. Table 4 presents the results as an unweighted macro-average across four tasks. If “Full” corresponds to the “anchor-residual” in Table 3, then its FD should be:

    (5.62+7.68+1.32+1.36)/4=3.995,(5.62+7.68+1.32+1.36)/4=3.995, rather than 1.74. Using the results for the corresponding four tasks in Table 2 yields 4.1825, which also fails to account for this value. Furthermore, the MAE/FD for PPG+ECG→ABP in Tables 2 and 3 are 0.54/8.18 and 0.58/7.68, respectively. Different training runs, checkpoints, or model configurations may account for these discrepancies, but the manuscript does not provide sufficient corresponding explanations, and the definitions of CardioPRF-F and CardioPRF-M are incomplete. My concern is whether these values collectively support the conclusions regarding the full model and ablation studies, rather than merely requiring superficial numerical consistency.

  2. The evaluation definitions are insufficient to explain the Min-MAE for deterministic methods.Section 4.1 defines Min-MAE as the best paired error among generated samples, but in Table 3, all anchor-only cases show Min-MAE values smaller than MAE—for example, 0.44 versus 0.49. For deterministic models with fixed inputs and no randomness, repeated generation should yield the same results, and the two metrics should align. Therefore, it is necessary to clarify whether random inference, aligned search, or different aggregation methods were used, and whether all methods employed comparable sampling budgets. The paper has labeled FD as the “raw Fréchet distance,” but the sample vectors, normalization, and statistical aggregation methods still need to be specified. Since the primary advantages center on FD, these definitions are critical for interpreting the results.

  3. Current evidence is insufficient to distinguish between population distribution matching and recovery from conditional variations. Table 3 shows that residual flow reduces FD while increasing pairing error on some tasks. This may represent a reasonable generative trade-off; the issue is that population FD alone cannot prove that the introduced variations are compatible with the given observations. There is currently a lack of direct evidence linking conditional consistency to sample-level variations. External evaluations do not yet fill this gap: for example, WESAD’s Macro-F1 of 0.31 lacks sufficiently clear specifications regarding tasks, downstream models, and participant groupings, making it difficult to assess its practical significance. I do not lower the score simply because the model does not lead on all metrics; rather, I focus on whether these results are sufficient to support the paper’s interpretation of conditional variations and the retention of physiological information.

Questions for authors
  1. Please describe the model variants, checkpoints, and evaluation configurations in Tables 2–4; provide the results for the four tasks used in Table 4; and explain how the FD of 1.74 for the “Full” configuration was calculated.
  2. Please clarify the calculation methods for MAE, Min-MAE, KS, and raw FD, as well as the number of samples for each condition, and explain why the two MAE metrics differ for the “deterministic anchor-only” configuration.
  3. Building on the clarification of the existing WESAD evaluation protocol, is it possible to use the available generated results to supplement a limited-scope consistency analysis—for example, comparing the consistency of rhythms generated multiple times under the same conditions versus variations across samples? If the current data are insufficient to support this interpretation, the conclusions may be narrowed accordingly.
2 comments
AuthorsOpenReview ↗

Response to Reviewer zLwV (Part 1/2)

We thank the reviewer for the detailed and constructive assessment, particularly for identifying inconsistencies in the reported results and ambiguities in the evaluation protocol. We have carefully revisited the experimental records and substantially revised the manuscript. We have also added cardiac-rhythm consistency experiments and five-seed inference evaluations to strengthen the empirical evidence.

1. Numerical consistency and model configurations

We acknowledge the numerical inconsistencies identified across Tables 2–4 of the original submission. After re-examining the evaluation records, we traced the discrepancies in two dual-observation settings to the inadvertent mixing of results from different checkpoints during table preparation. We have corrected these entries using the corresponding validation-selected checkpoints, without selecting checkpoints based on test-set performance.

Specifically, for PPG+ABP→ECG, the reported MAE/raw FD values have been corrected from 0.28/1.61 to 0.29/1.36. For PPG+ECG→ABP, they have been corrected from 0.54/8.18 to 0.58/7.68.

We also corrected the aggregate raw FD reported in the original Table 4. As the reviewer noted, the four corresponding task-level values (5.62, 7.68, 1.32, and 1.36) yield an unweighted mean of 3.995, rather than the previously reported 1.74. In the revised manuscript, Table 4 presents the results separately for each task, avoiding aggregation of raw FD across target modalities with different waveform scales. The associated descriptions have been updated accordingly.

We have further clarified the model configurations in Section 4.1 and Appendix B. CardioPRF is the unified model trained on all nine directed tasks; CardioPRF-F is trained specifically for an individual translation task; and CardioPRF-M uses the same unified CardioPRF model with two observed modalities at inference. Appendix E (Tables 7–9) provides the complete nine-task results and extended ablations.

These corrections preserve the central empirical finding: anchor–residual generation achieves lower raw FD than direct conditional flow across the representative tasks, while maintaining competitive paired reconstruction accuracy.

2. Evaluation metrics and sampling protocol

We appreciate the reviewer pointing out the ambiguity in the original Min-MAE description. We have corrected Section 4.1 and provided the precise metric definitions and aggregation procedures in Appendix D (Eqs. 11–15).

For the main waveform-generation evaluation, one waveform is generated per conditioning window. Standard MAE measures the pointwise difference from the paired recorded target without temporal alignment. Min-MAE follows the bounded temporal-shift evaluation used in UniCardio: for each 512-sample waveform at 128 Hz, we search integer shifts within ±128 samples (±1 second) and select the minimum MAE over the valid, non-circular overlapping region. The per-window minima are then averaged over the evaluation set.

Importantly, the 257 candidate shifts represent temporal alignments, not multiple stochastic generations. Thus, Min-MAE can be smaller than MAE even for deterministic anchor-only predictions. The same alignment procedure is applied to all evaluated methods.

For distributional evaluation, KS is computed from the empirical amplitude distributions of each generated–reference waveform pair and then averaged across windows. Raw FD is calculated using the empirical means and covariance matrices of the complete collections of 512-dimensional waveform vectors, without a pretrained feature extractor. All methods are evaluated on matched test windows using consistent metric implementations.

For the five-seed inference analysis in Appendix E (Tables 10–11), each seed produces a separate full-test-set evaluation using fixed model checkpoints. The four metrics are computed independently for each run, and the resulting values are summarized by their mean and sample standard deviation.

AuthorsOpenReview ↗

Response to Reviewer zLwV (Part 2/2)

3. Additional evidence for conditional rhythm consistency and sampling stability

We appreciate the reviewer's distinction between matching the overall waveform distribution and preserving physiological characteristics associated with individual observations. To provide evidence beyond marginal distributional similarity, we introduced a cardiac-rhythm consistency experiment in Section 4.4 (Table 5) and Appendix F.

On the subject-disjoint MIMIC-BP test set, we compare anchor-only prediction, direct conditional flow, and anchor–residual generation using identical PPG observations and paired ECG references. We evaluate cardiac-period errors relative to both the observed PPG and recorded ECG, together with R-peak matching against the recorded ECG.

Anchor–residual generation achieves the lowest cardiac-period MAE relative to the observed PPG (54.93 ms, compared with 58.30 and 80.86 ms) and the recorded ECG (60.37 ms, compared with 62.28 and 72.37 ms). Its R-peak micro-F1 also improves to 0.566, compared with 0.512 for anchor-only prediction and 0.498 for direct flow.

These results provide direct, rhythm-specific evidence that generated ECG better retains cardiac timing information related to its conditioning PPG, complementing the waveform-level distributional measurements.

We additionally evaluated anchor–residual and direct flow using five inference seeds across all nine directed generation tasks, with fixed checkpoints (Appendix E, Tables 10–11). Anchor–residual flow achieves lower mean raw FD in every task. For example, PPG→ECG raw FD is 1.23 ± 0.07, compared with 5.55 ± 0.09 for direct flow; for PPG→ABP, the corresponding values are 5.91 ± 0.33 and 13.40 ± 0.39. Paired reconstruction errors remain competitive across the tasks.

Finally, we have strengthened the downstream evaluation with a frozen BP-regressor experiment (Section 4.5, Table 6; Appendix H). Following the downstream evaluation approach adopted in prior work such as UniCardio, we average five independently generated ECG realizations for each PPG input before substituting the resulting ECG into a fixed BP predictor. The recorded PPG and regressor parameters remain unchanged. On external Cuffless BP, CardioPRF-generated ECG yields SBP/DBP MAE of 17.70/8.01 mmHg, close to 17.79/7.90 mmHg obtained using recorded ECG. The WESAD evaluation is also retained in Appendix G.

Together, the revised results provide complementary evidence from waveform-distribution matching, inference stability, observation-related cardiac rhythm, and downstream prediction. We have also refined the discussion to distinguish rhythm-level physiological consistency from broader claims about the complete conditional waveform distribution.

We appreciate the reviewer's careful scrutiny, which has helped us improve the numerical accuracy, evaluation transparency, and empirical support of the manuscript.

Reviewer_X7ynOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines how missing cardiovascular signals, including electrocardiography (ECG), photoplethysmography (PPG), and arterial blood pressure (ABP), can be generated from available signals. The authors propose CardioPRF, a generative framework that separates predictable physiological patterns from uncertain variations. The method uses deterministic representations for predictable patterns and conditional flow matching to model the remaining variations, with both global and local information used to guide generation. Experiments on MIMIC-BP show that the proposed method generates signals with better distributional similarity than deterministic and direct-flow baselines. Evaluations on WESAD and Cuffless BP also show that the generated signals preserve useful physiological information for downstream prediction tasks.

Critical strengths
  • simple factorization and easy to understand
Critical weaknesses
  • paper feels very incomplete. authors should dedicate further analysis and wider exploration and follow the guidelines for an iclr paper
  • method itself resembles residual generative modeling and contributions are unclear
  • imporovements are not consistent across metrics
  • low marginal distributional discrepancy does not necessarily imply that generated signals are physiologically consistent with their conditioning observations
  • the ablation between direct flow and anchor-residual flow demonstrates improvements in distributional fidelity, but it remains unclear whether these improvements arise from the residual decomposition itself, differences in optimization difficulty, or the additional deterministic predictor.
  • no external validation
Questions for authors

see weakness and provide additional experiments to address weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer X7yn (Part 1/2)

We thank the reviewer for the critical assessment and constructive suggestions regarding methodological novelty, experimental completeness, and physiological consistency. We have substantially revised the manuscript to better articulate the proposed formulation and strengthen its empirical evaluation. The revision includes complete nine-task ablations, five-seed inference analysis, cardiac-rhythm consistency measurements, and expanded cross-dataset downstream evaluation.

1. Methodological contribution and relation to residual generative modeling

We agree that residual modeling is an established concept, and our contribution is not the introduction of additive decomposition itself. Instead, CardioPRF develops a predictable–residual formulation for underdetermined cardiovascular generation, explicitly separating observation-supported target prediction from stochastic modeling of the remaining variation.

The key distinction from direct conditional generation lies in how these two components are learned and coordinated. CardioPRF first predicts an observation-conditioned, target-specific latent anchor, using global cross-modal context and locally resolved temporal evidence. Conditional flow matching then models the residual relative to this anchor, rather than generating the complete target latent through a single stochastic process.

Importantly, the factorization is established through stage-wise optimization, rather than an unconstrained decomposition. The target autoencoders and deterministic anchors are trained before residual flow; during flow training, stop-gradient operations separate the residual target and the anchor directly supplied to the velocity network. Subsequent joint refinement adapts the complete framework.

This design supports nine directed generation tasks with different observed subsets and target modalities within a shared model. We have expanded Sections 2–3 to clarify these architectural and optimization choices and their relationship to existing cardiovascular generators and residual-based methods, including ResShift.

The contribution therefore lies in the integrated modeling and optimization of a target-specific predictable reference and its conditional residual, rather than in residual decomposition considered in isolation.

2. Extended ablations and distributional improvements

We appreciate the concern that improvements in the original submission were not uniform across all metrics. The revised manuscript now more explicitly distinguishes paired reconstruction accuracy from waveform-distribution fidelity, presenting the corresponding results at the task level.

The generation-formulation ablation (Table 3) compares deterministic anchor-only prediction, direct conditional flow, and anchor–residual flow under four representative configurations. Anchor-only prediction achieves the lowest paired errors, whereas direct flow considerably improves distributional matching. Anchor–residual flow further reduces raw FD across all four tasks and improves paired reconstruction over direct flow for both ECG generation settings.

For PPG→ECG, anchor–residual flow achieves MAE 0.31 and raw FD 1.32, compared with 0.33 and 5.50 for direct flow. For PPG→ABP, raw FD decreases from 12.94 to 5.62, with comparable paired MAE (0.59 versus 0.60).

We have extended these comparisons to all nine tasks in Appendix E (Tables 7–9), including global/local conditioning and generation-formulation ablations. Revised Table 4 also reports conditioning-pathway ablations separately for each representative task, revealing their different contributions across ECG and ABP targets.

We further conducted five-seed inference evaluations with fixed checkpoints across all nine tasks (Appendix E, Tables 10–11). Anchor–residual flow retains lower mean raw FD in every task. For PPG→ECG, raw FD is 1.23 ± 0.07, compared with 5.55 ± 0.09 for direct flow; for PPG→ABP, it is 5.91 ± 0.33, compared with 13.40 ± 0.39.

These results demonstrate consistent distributional improvements across observation settings and inference-noise realizations, with competitive reconstruction performance. The comparisons support the effectiveness of the complete factorized formulation, including its learned anchor, residual generation, and optimization strategy, rather than attributing the improvements to the algebraic decomposition alone.

AuthorsOpenReview ↗

Response to Reviewer X7yn (Part 2/2)

3. Direct evaluation of conditioning-related cardiac rhythm

We agree that marginal waveform-distribution similarity alone does not establish whether generated signals preserve physiological characteristics associated with their conditioning observations. We have therefore introduced a cardiac-rhythm consistency experiment on the subject-disjoint MIMIC-BP test set (Section 4.4, Table 5; Appendix F).

Using identical PPG observations and paired recorded ECG references, we compare anchor-only prediction, direct flow, and anchor–residual generation. Beyond raw FD, we measure cardiac-period agreement with both the observed PPG and recorded ECG, together with R-peak matching against the recorded ECG.

Anchor–residual generation achieves the lowest cardiac-period error relative to the conditioning PPG (54.93 ms, versus 58.30 and 80.86 ms) and the recorded ECG (60.37 ms, versus 62.28 and 72.37 ms). It also achieves the highest R-peak micro-F1 of 0.566, compared with 0.512 for anchor-only prediction and 0.498 for direct flow.

These findings provide additional physiological evidence beyond distributional matching: the proposed formulation better preserves cardiac timing information related to the observed signal and improves beat-level agreement with the recorded target. We have incorporated this analysis into the main manuscript and refined the interpretation to distinguish rhythm-level consistency from overall waveform-distribution fidelity.

4. Cross-dataset evaluation and downstream utility

We would also like to clarify that the original submission already included downstream evaluations on WESAD and Cuffless BP, neither of which was used for generator training or checkpoint selection. These experiments examine whether generated signals retain useful physiological information under dataset shift.

We have strengthened this evaluation through an additional frozen BP-regressor experiment (Section 4.5, Table 6; Appendix H). A dual-branch regressor is trained on recorded MIMIC-BP PPG and ECG, then evaluated with its ECG input replaced by synthesized ECG while keeping the observed PPG and regressor parameters unchanged.

Following the downstream evaluation approach adopted in prior cardiovascular generation work, including UniCardio, five generated ECG realizations are averaged for each input before being supplied to the fixed regressor. The same evaluation procedure is applied to the compared generation methods.

On external Cuffless BP, CardioPRF-generated ECG achieves SBP/DBP MAE of 17.70/8.01 mmHg, close to 17.79/7.90 mmHg obtained using recorded ECG. On the MIMIC-BP test set, it achieves 13.15/8.56 mmHg, compared with 13.22/8.66 mmHg using recorded ECG. These results show that synthesized ECG retains useful information for downstream BP estimation under the evaluated protocol.

We have also retained the external WESAD evaluation in Appendix G, where CardioPRF achieves the highest Macro-F1 (0.31) among the compared generation methods. Appendix J additionally reports BP estimation results obtained directly from generated ABP waveforms on Cuffless BP.

Together, these evaluations provide evidence of downstream utility across datasets and prediction settings, complementing the waveform-generation and rhythm-consistency results.

We believe the revised manuscript now presents a clearer account of the methodological contribution and more comprehensive empirical evidence. The nine-task ablations, inference stability, cardiac-rhythm analysis, and cross-dataset evaluations strengthen the original findings while clarifying their scope. We appreciate the reviewer's comments, which helped motivate these improvements.

Reviewer_X7ynOpenReview ↗

Follow up comments

Can the authors resubmit their manuscript with tracked changes (via a different color of text). It is hard to distinguish what was added. I think the expeirmentation based on the review is much stronger but the new cuffless BP experiments dont provide any useful signal. A MAE of ~18/8 mmHg does not convince me that it is a sound method despite it being the leading model.

AuthorsOpenReview ↗

Response to Reviewer X7yn (Follow-up)

We thank the reviewer for the follow-up comments and for acknowledging that the additional experiments have substantially strengthened the evaluation. We appreciate both concerns and address them below.

1. Highlighted manuscript revisions

Following the reviewer's suggestion, we have uploaded a color-highlighted version of the revised manuscript, with substantive additions and modifications marked in blue to facilitate comparison with the original submission.

The main changes include clarified model design and stage-wise optimization (Section 3), more precise evaluation protocols (Section 4.1 and Appendix D), complete nine-task ablations and five-seed inference analysis (Section 4.3 and Appendix E), new cardiac-rhythm consistency experiments (Section 4.4 and Appendix F), and expanded downstream evaluations (Section 4.5 and Appendices H–J).

2. Experimental reliability and interpretation of the Cuffless BP results

We appreciate the reviewer's concern regarding the absolute BP estimation errors on Cuffless BP. We agree that an SBP MAE of approximately 18 mmHg does not establish clinically accurate BP estimation, and we do not intend to make such a claim.

To clarify the reliability of our experimental protocol, all source physiological recordings and reference BP measurements are derived from real recorded datasets. Our main experiments use strictly subject-disjoint training, validation, and test partitions of MIMIC-BP (971/176/204 subjects), with consistent preprocessing and signal-quality screening. Model checkpoints are selected based on validation performance rather than test results. The external Cuffless BP dataset is not used for generator training or checkpoint selection.

The purpose of the downstream BP experiment is to examine whether generated ECG can serve as a useful auxiliary representation for BP estimation and how it performs relative to recorded ECG under a controlled evaluation protocol.

Specifically, we train a dual-branch ConvLSTM regressor using recorded PPG and ECG from MIMIC-BP. During evaluation, the regressor remains frozen, while recorded ECG is replaced by generated ECG. The observed PPG, regression targets, and evaluation windows remain unchanged. CardioPRF uses the same validation-selected generator checkpoint as in the main waveform-generation experiments.

On external Cuffless BP, the frozen regressor achieves 17.79/7.90 mmHg (SBP/DBP MAE) using recorded ECG, compared with 17.70/8.01 mmHg using CardioPRF-generated ECG. Moreover, the independently trained PPG-only reference yields substantially higher errors of 25.67/9.25 mmHg.

Although the PPG-only predictor is trained separately, these results suggest that CardioPRF-generated ECG can provide a useful complementary representation for downstream BP estimation, while achieving performance closely comparable to that obtained with recorded ECG under the same frozen-regressor protocol.

We emphasize that the value of this experiment lies in demonstrating the downstream utility of generated ECG under cross-dataset evaluation, rather than establishing clinically accurate BP estimation. The remaining absolute errors reflect the challenges of BP prediction under dataset shift. We also acknowledge that recovering absolute pressure levels directly from generated ABP remains challenging, as discussed in Appendix J.

Importantly, BP regression is not the primary objective of CardioPRF. The main contribution concerns conditional cardiovascular waveform generation, supported by complete nine-task comparisons, generation-formulation and conditioning-pathway ablations, five-seed inference evaluations, and cardiac-rhythm consistency measurements. In particular, the new rhythm analysis (Table 5) provides complementary evidence of conditioning-related cardiac timing, independently of the BP regression results.

We sincerely appreciate the reviewer's prompt and constructive follow-up, as well as their continued engagement with our work. These valuable comments have helped us improve the presentation of our experimental evidence and more clearly delineate the scope of our conclusions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.