acceptodds
Withdrawn from ICLR 2027

Physiology-Guided Age-Stage-Consistent Diffusion for Sleep EEG Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Conditional diffusion models can generate physiological time series that appear realistic in spectral and feature statistics; however, marginal realism does not establish whether age-related changes remain compatible with an assigned sleep stage. We study this globally realistic yet conditionally contradictory behavior as age–stage physiological inconsistency in sleep EEG. To address it, we propose Phy-DiffKAN, which represents condition-dependent physiological structure as a factorized prior field over an interpretable 12-dimensional coordinate space. The prior field decomposes age, recorded sex, sleep stage, and their pairwise interactions into six branches, smoothly approximates the finite-anchor states of each branch, and projects them into physiology-structured tokens for a Transformer diffusion denoiser. Branch-drop and branch-mismatch constraints further enforce branch necessity and value specificity. We evaluate Phy-DiffKAN on a subject-disjoint Sleep-EDF cohort with controlled demographic allocation and complete per-subject stage coverage, using overall generation quality, fixed-stage preservation, and age-edit responsiveness. Phy-DiffKAN achieves an FID of 0.53, a feature-manifold F1 score of 0.78, a LogPSD distance of 0.06, and a five-stage balanced accuracy of 0.87. Under youngold REM reassignment with fixed diffusion noise, Phy-DiffKAN preserves REM-stage identity, suppresses leakage into N2, and retains near-real age decodability; complementary N1 age sweeps show the same stage-preservation pattern beyond REM.

Primary area:
generative models
Keywords:
Generative models, Diffusion Models, Sleep EEG signals

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

This paper proposes Phy-DiffKAN to tackle age–stage physiological inconsistency in conditional sleep EEG generation. Phy-DiffKAN constructs a factorized physiological prior field comprising 12 interpretable coordinates and organizes age, recorded sex, sleep stage, and their pairwise interactions into six condition branches. Under the strict setting that all models are evaluated on the same subject-disjoint Sleep-EDF cohort with equal subject allocation across age strata, near-balanced recorded sex, and complete per-subject stage coverage, the Phy-DiffKAN achieved promising results.

Critical strengths

S1: The authors formulate age–stage physiological inconsistency in sleep EEG, which I found to be an interesting and potentially useful setting.

S2: The results are meaningful, revealing that similar overall generation quality can correspond to substantially different age–stage behaviors.

Critical weaknesses

W1: The experiment is limited to a small dataset. I would like to see how it performs on larger datasets. Sleep EEG is not rare, so testing only on a small number of subjects from a single dataset is questionable.

W2: I found it hard to understand the motivation for generating sleep EEG, because sleep EEG is easy to collect. And there are many sleep EEG datasets out there.

Questions for authors

What is the motivation to generate sleep EEG data?

What is the model performance for some large-scale sleep EEG datasets, like HMC and Dreem datasets?

Does the model performance largely depend on the number of channels in sleep EEG, as it may vary from one to many?

1 comment
AuthorsOpenReview ↗

Author Response to Reviewer Gike

Author Response to Reviewer Gike

We sincerely thank the reviewer for the constructive feedback and for recognizing the value of the age–stage inconsistency problem and our physiology-guided conditioning framework.

1. Dataset Scale and Generalization (W1)

We acknowledge that the original evaluation uses only 30 subjects from Sleep-EDF. To further investigate generalizability, we plan to extend our experiments to the HMC dataset and evaluate the proposed framework on an additional sleep EEG cohort.

2. Motivation for Sleep EEG Generation (W2 and Q1)

Our goal is not to replace conventional EEG acquisition, but to investigate controllable generation of physiologically conditioned sleep EEG. By specifying age, recorded sex, and sleep stage, our framework aims to generate signals under controlled conditions. Potential applications include controlled data augmentation, systematic evaluation of conditional generation models, and analysis of demographic-related physiological variations. We will clarify these motivations and potential applications in the revised manuscript.

3. Evaluation on HMC (Q2)

We plan to conduct additional experiments on the HMC dataset to evaluate the applicability of Phy-DiffKAN beyond the original Sleep-EDF setting. The experimental cohort and evaluation protocol will be determined according to the available demographic annotations and sleep-stage labels.

4. Influence of EEG Channel Number (Q3)

We plan to investigate different EEG channel configurations on HMC, including two-, three-, and four-channel settings where supported by the available recordings. The two-channel configuration will serve as a reference to distinguish dataset-related differences from the effects of channel number.

We sincerely appreciate the reviewer's suggestions. These additional evaluations will help clarify the applicability and limitations of Phy-DiffKAN beyond the original Sleep-EDF setting.

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

This paper investigates age–stage inconsistency in conditional sleep EEG generation, where age edits can shift generated signals away from the requested sleep stage despite plausible overall quality. It proposes Phy-DiffKAN, which represents a predefined physiological prior field using six branches for age, recorded sex, sleep stage, and their pairwise interactions in a shared 12-dimensional coordinate system. A KAN encoder is distilled from finite prior anchors and then frozen, with its branch outputs projected into separate conditioning tokens for a diffusion Transformer. Branch-drop and branch-mismatch objectives encourage branch necessity and value specificity. Evaluation uses a subject-disjoint Sleep-EDF Expanded cohort with 21 training and nine test subjects, assessing generation quality, stage preservation under fixed-noise age edits, and age-group decodability within REM. The reported comparisons show improved REM preservation and reduced REM-to-N2 leakage while retaining age information, complemented by N1 age sweeps, classifier-independent waveform-marker analyses, and ablations.

Critical strengths
  1. A valuable diagnostic perspective. The paper demonstrates that marginally realistic sleep EEG can still violate age–stage conditioning. Its fixed-stage, matched-noise age-edit tests expose conditional control failures that aggregate generation metrics may overlook, motivating more informative evaluation of physiological signal generators.

  2. Promising empirical evidence with complementary validation. Phy-DiffKAN achieves an old-REM recall of 0.75, compared with 0.19–0.34 for external baselines, while retaining an age-probe balanced accuracy of 0.72, close to the real-data reference of 0.73. The subject-disjoint evaluation, complementary N1 stress tests, waveform-marker analyses, and ablations provide converging evidence that the method improves stage preservation without simply discarding age information.

Critical weaknesses
  1. The core consistency claim is not directly validated at the sample level. The paper evaluates stage preservation and age decodability separately, but does not report how often the same generated samples satisfy both conditions. There is also a potential issue with the age probe: it is trained on real REM epochs but evaluated on baseline outputs that may be classified as N2. This stage mismatch could affect age-decoding performance, making it harder to determine whether the model truly preserves both age and stage information in the same sample.

  2. The contribution of physiological priors is not fully established. The existing ablations show that the proposed design performs better than several alternatives, but they do not completely isolate the role of physiological knowledge. A capacity-matched six-branch model trained without physiological prior supervision, while keeping the same diffusion architecture and branch-specific losses, would help test whether the priors themselves improve performance. A separate direct-rule conditioning baseline would clarify whether learning the prior field with KAN offers an advantage over using the rules directly. The waveform-marker analyses provide useful supporting evidence, but do not by themselves establish that the learned age effects are physiologically correct.

Questions for authors
  1. Joint consistency — clarification and analysis of existing samples: Do Tables 3 and 4 use the same generated samples and age assignments? If so, what fraction of matched-noise age-edit pairs simultaneously preserve the requested sleep stage and yield the correct target-age-group classification? If not, can this joint statistic be computed from the existing samples?

  2. Age-probe validity — clarification and targeted analysis: How should age-decoding performance be interpreted when baseline outputs shift away from REM, given that the age probe was trained exclusively on real REM epochs? Does the comparison change when the probe is evaluated on generated samples classified as REM, and what fraction of each method's samples is retained under this restriction?

  3. Contribution attribution — targeted validation: What evidence isolates the contribution of physiological prior supervision from that of the six-branch architecture and branch-specific losses? Would the improvement persist against a capacity-matched six-branch conditioner trained without physiological prior supervision, while retaining the same diffusion architecture and branch-specific losses?

  4. Reporting completeness and reproducibility — clarification: Several implementation details remain insufficiently specified, including the preprocessing and normalization procedures, key denoiser configurations, and adaptations used to provide demographic conditions to external baselines. The age encoding also differs between the main text (a/90a/90) and Appendix A.5 (a/100a/100). Could the authors clarify the actual implementation and provide the configurations needed to assess whether the comparisons are reproducible and appropriately controlled?

8 comments

Author Response to Reviewer dsyD

We sincerely thank the reviewer for the careful evaluation and constructive feedback. We appreciate the recognition of the age–stage inconsistency problem and the potential value of our physiology-guided conditioning framework.

After carefully considering the reviewer's comments, we recognize that several aspects of our work require further validation, particularly the sensitivity and physiological validity of the handcrafted priors, the contribution of physiological prior supervision, and the simultaneous preservation of age and sleep-stage information in generated EEG.

To address these concerns, we plan to conduct additional controlled experiments and extend our physiological analyses. Specifically, our revisions will focus on: (1) sample-level joint age–stage consistency evaluation; (2) stage-restricted and five-stage age-decoding experiments; (3) controlled ablation studies to isolate the contribution of physiological prior supervision; (4) direct analysis of age-dependent physiological characteristics in generated EEG; and (5) sensitivity and robustness analyses of the physiological priors and training objectives.

We provide our detailed responses below.


W1. Limited Justification of the Handcrafted Physiological Prior

We appreciate the reviewer's concern regarding the manually specified physiological anchors, interaction coefficients, and branch weights.

We would like to clarify that the 12 physiological coordinates are designed as interpretable, literature-informed prior representations. Their numerical values encode the expected direction and relative strength of physiological modulation on a shared dimensionless scale, rather than representing directly measured EEG power, amplitude, or event rates.

Nevertheless, we agree that the sensitivity of these manually specified values and their physiological implications require further investigation.

To address this concern, we plan to conduct systematic sensitivity analyses from three perspectives.

First, we will perturb the numerical values of the 12 physiological anchors individually and jointly at different magnitudes, and examine their effects on prior-fitting errors and the learned physiological response functions.

Second, we will investigate the sensitivity of the age-related and pairwise interaction coefficients, including age–stage, age–sex, and sex–stage interactions, to determine whether the learned representations depend excessively on specific manually assigned coefficients.

Third, we will examine the effects of varying the six branch weights and the loss weights used in prior-field distillation and diffusion training, including the branch-mismatch and branch-drop objectives.

We will report prior-fitting errors, physiological constraint violations, and the stability of the learned response curves under these settings.

Importantly, we will distinguish the numerical robustness of the prior encoder from the physiological validity of generated EEG. The latter requires direct evaluation against real EEG, which we will address through the additional waveform-level physiological analyses described below.


W2. Insufficient Isolation of the Methodological Contributions

We agree that the contributions of physiological prior supervision, factorized conditioning, and KAN parameterization should be more explicitly distinguished.

In the original submission, the Phy_MLP ablation primarily compares KAN and MLP encoders under the same physiological targets and six-branch conditioning framework. However, this comparison alone does not fully establish the contribution of physiological prior supervision itself.

To address this limitation, we plan to introduce a capacity-matched six-branch model without physiological prior supervision. This model will retain the same diffusion architecture, six-branch conditioning structure, and branch-specific training objectives, while replacing the physiological prior supervision with learnable condition representations.

This controlled comparison will allow us to investigate whether the proposed physiological prior provides additional benefits beyond the six-branch architecture and branch-sensitive objectives.

Furthermore, we will clarify the role of the existing Phy_MLP comparison and examine the prior-fitting accuracy, model capacity, optimization configurations, and hyperparameter settings of Phy_MLP and Phy-DiffKAN to ensure a fair comparison.

We will also investigate a direct-rule conditioning baseline, in which the predefined physiological coordinates are directly provided to the diffusion model without KAN-based prior-field approximation. This additional comparison will help clarify the benefit of learning a continuous physiological prior field through KAN.

AuthorsOpenReview ↗

Question 1 and 2


Q1. Joint Consistency: Clarification and Analysis of Existing Samples

We appreciate the reviewer for identifying an important limitation in our original evaluation.

We agree that separately evaluating sleep-stage preservation and age decodability does not directly demonstrate that both properties are simultaneously retained within the same generated EEG sample.

To address this concern, we plan to conduct a sample-level joint consistency analysis. Specifically, the same generated EEG samples will be evaluated using both a frozen five-stage sleep classifier and an age classifier.

For each sample, we will determine whether its predicted sleep stage matches the requested stage and whether its predicted age group matches the assigned age condition.

We will report the joint consistency accuracy, defined as the proportion of generated samples that simultaneously satisfy both requirements:

J=Nstage correct and age correctNallJ=\frac{N_{\text{stage correct and age correct}}}{N_{\text{all}}}

We will additionally report the four possible outcomes: both conditions correctly satisfied, only stage correctly preserved, only age correctly decoded, and neither condition correctly satisfied.

This analysis will provide a more direct evaluation of the central claim of our work, namely, that the proposed framework preserves age-related information while maintaining the requested sleep-stage identity.



Q2. Age-Probe Validity: Clarification and Targeted Analysis

We acknowledge the reviewer's concern regarding the potential distribution mismatch introduced by the REM-only age probe.

Since the original age classifier is trained exclusively on real REM EEG, applying it to generated samples that may have shifted toward other sleep stages could affect the reliability of age-decoding results.

To address this issue, we plan to conduct a stage-restricted age-decoding analysis. Specifically, we will first identify generated samples classified as the requested REM stage using the frozen sleep-stage classifier. We will then evaluate age decodability only on these retained REM samples using the REM-trained age probe.

We will report both the resulting age-decoding performance and the corresponding sample-retention ratio for each method.

Furthermore, we plan to extend the age-decoding evaluation to all five sleep stages. Specifically, five separate age probes will be trained using real EEG from Wake, N1, N2, N3, and REM, respectively, with subject-disjoint training and testing splits.

Each stage-specific age probe will then be used to evaluate generated EEG samples corresponding to the same requested stage. This design will help reduce potential confounding effects caused by differences in sleep-stage distributions.

Results of Q1 & Q2

We have now performed the requested sample-level joint consistency analysis. For 500 generated 25-year REM samples and 500 generated 65-year REM samples, the same samples were evaluated by the frozen five-stage sleep classifier and the age classifier. 546/1000 (54.6%) satisfied both the requested sleep stage and age condition; 143 (14.3%) satisfied only the sleep-stage condition; 246 (24.6%) satisfied only the age condition; and 65 (6.5%) satisfied neither. Marginal accuracy was 68.9% for sleep-stage preservation and 79.2% for age decodability. The joint consistency accuracy was 54.6% (95% CI, approximately 51.5–57.7%), which is essentially equal to the product of the two marginal accuracies (68.9% × 79.2% = 54.6%), indicating no evident negative trade-off between stage preservation and age decodability in this sample.

AuthorsOpenReview ↗

Question 3 and additional Validation; Question4

Q3. Contribution Attribution: Targeted Validation

We agree that additional controlled experiments are necessary to isolate the contribution of physiological prior supervision.

As described in our response to W2, the proposed capacity-matched six-branch model without physiological prior supervision will retain the original diffusion architecture and branch-specific objectives while removing the predefined physiological supervision.

We will compare this model with Phy-DiffKAN using generation quality, sleep-stage preservation, age decodability, and joint age–stage consistency metrics.

In addition, the direct-rule conditioning baseline and the existing Phy_MLP variant will provide complementary comparisons to distinguish the effects of physiological prior specification, KAN-based approximation, and encoder architecture.

Together, these experiments will provide a more systematic evaluation of the contribution of physiological prior supervision and its interaction with the proposed factorized conditioning design.


Additional Validation: Age-Dependent Physiological Characteristics

We agree that age decoding alone does not fully establish whether the generated EEG exhibits appropriate age-dependent physiological changes.

To provide more direct evidence, we plan to extend the existing waveform-marker analysis to age-related physiological characteristics.

Specifically, we will fix the recorded sex and requested sleep stage while varying the assigned age. We will generate EEG under different age conditions and extract the corresponding physiological features from the generated waveforms, including delta, theta, alpha, sigma, and beta power, slow-wave activity, spindle density, spindle amplitude, spindle frequency, K-complex characteristics, mixed-frequency activity, and spectral slope.

We will then compare the age-dependent trends in generated EEG with those observed in real EEG, focusing on the direction and magnitude of feature changes. These analyses will be conducted separately across the five sleep stages.

This evaluation will complement classifier-based age decoding by directly examining whether the age-related modulation encoded by the physiological prior is reflected in the generated waveform characteristics.


Additional Validation: Generalization Across Datasets and Subject Splits

We acknowledge that the original evaluation is based on a relatively small cohort from a single dataset and one subject-disjoint split.

To further investigate the generalizability of the proposed framework, we plan to extend the evaluation to the HMC dataset using different EEG channel configurations, including three- and four-channel settings where supported by the available recordings.

We will document the channel selection, preprocessing procedures, demographic distributions, and subject-disjoint evaluation protocols.

We will also investigate the robustness of the proposed method across multiple subject-disjoint splits, while maintaining demographic allocation and sleep-stage coverage as consistently as possible.

These additional experiments will help assess whether the observed age–stage consistency is dependent on a particular subject allocation or recording configuration.


Q4. Reporting Completeness and Reproducibility

We thank the reviewer for identifying several implementation details that require clarification.

We will revise the manuscript to provide a more complete description of EEG preprocessing, normalization, channel selection, denoiser configurations, and the demographic-conditioning procedures used for external baselines.

We will also carefully verify the actual age-normalization implementation and correct the inconsistency between the main text and Appendix A.5. The age-normalization formula will be reported consistently throughout the manuscript according to the actual implementation.

Furthermore, we will provide additional details regarding the optimization configurations and training procedures of the compared models to improve the transparency and reproducibility of the experimental comparisons.

Reviewer_4UdbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Phy-DiffKAN to address age–stage inconsistency in conditional sleep EEG generation. It combines a factorized physiological prior, KAN-based condition encoding, and branch-sensitive objectives within a diffusion Transformer. Experiments on Sleep-EDF demonstrate improved signal fidelity, sleep-stage preservation, and age decodability over existing generators.

Critical strengths

S1. Meaningful problem formulation. The paper identifies age–stage inconsistency as an important limitation of conditional EEG generation, highlighting that realistic signals may not preserve the intended physiological conditions. S2. Structured physiology-guided conditioning. The factorized prior explicitly models demographic and sleep-stage effects and their interactions, providing a structured and interpretable alternative to shared condition embeddings. S3. Strong empirical performance. The method consistently outperforms existing generators across signal quality and conditional consistency metrics, with ablation studies supporting several design choices.

Critical weaknesses

W1. Limited justification of the handcrafted physiological prior. The numerical anchors, interaction coefficients, and branch weights are manually specified without quantitative calibration, leaving their physiological validity and sensitivity insufficiently established. W2. Insufficient isolation of the methodological contributions. Although several ablations are provided, it remains unclear whether the improvements arise from physiological knowledge, factorized conditioning, or the KAN parameterization itself. In particular, the substantial performance gap between Phy_MLP and Phy-DiffKAN warrants further investigation. W3. Limited validation of physiological consistency and generalization. The main age–stage consistency evaluation relies on pretrained classifiers rather than directly validating age-dependent physiological changes. Furthermore, experiments use only 30 subjects from one dataset with a single subject-disjoint split, limiting evidence of generalization.

Questions for authors

Q1. How sensitive is Phy-DiffKAN to the manually specified physiological anchors, interaction coefficients, and branch weights? Q2. Can the authors clarify whether Phy_MLP achieves comparable prior-fitting accuracy and receives equivalent optimization and hyperparameter tuning? Q3. Can the authors directly validate age-dependent physiological changes after age reassignment and assess robustness across different subject splits?

4 comments
AuthorsOpenReview ↗

Author Response to Reviewer 4Udb

Author Response to Reviewer 4Udb

We sincerely thank the reviewer for the careful evaluation and constructive feedback. We appreciate the recognition of the significance of the age–stage inconsistency problem, the structured physiology-guided conditioning framework, and the empirical performance of Phy-DiffKAN.

We understand that the main concerns relate to the sensitivity of the handcrafted physiological priors, the attribution of improvements to different methodological components, and the limited validation of physiological consistency and generalization.

To address these concerns, we plan to prioritize three aspects in the revision: (1) systematic sensitivity analyses of physiological anchors, interaction coefficients, branch weights, and training loss weights; (2) a controlled comparison of the prior-fitting capabilities and optimization configurations of Phy_MLP and Phy-DiffKAN; and (3) further analysis of age-dependent physiological characteristics and the robustness of the evaluation.

W1. Limited Justification of the Handcrafted Physiological Prior

We appreciate the reviewer's concern regarding the manually specified physiological anchors, interaction coefficients, and branch weights.

Our 12-dimensional physiological coordinates are designed as interpretable, literature-informed prior representations. Their numerical values encode the expected directions and relative strengths of physiological modulation associated with age, recorded sex, and sleep stage, rather than representing directly measured physiological quantities.

We agree that their sensitivity requires more systematic investigation. To address this concern, we plan to conduct controlled sensitivity analyses covering three aspects.

First, we will perturb the 12 physiological anchor values individually and jointly at different magnitudes, and examine their effects on prior-fitting errors and learned physiological response functions.

Second, we will investigate the sensitivity of the pairwise interaction coefficients, including age–sex, age–stage, and sex–stage interactions.

Third, we will vary the six branch weights and the loss weights used in prior-field distillation, including the anchor, ranking, monotonicity, and orthogonality objectives.

We will report the corresponding fitting errors, physiological constraint violations, and changes in the learned response curves. These analyses will provide quantitative evidence regarding the numerical stability of the proposed physiological prior representation.

W2. Insufficient Isolation of the Methodological Contributions

We agree that the contributions of physiological prior supervision and KAN parameterization deserve further clarification.

In the original submission, the Phy_MLP ablation primarily compares KAN and MLP encoders under the same physiological targets and six-branch conditioning framework. However, we recognize that the substantial performance gap between these variants requires a more controlled analysis.

To address this concern, we plan to conduct a systematic comparison of Phy_MLP and Phy-DiffKAN in terms of prior-fitting accuracy, model capacity, optimization configurations, and hyperparameter settings.

Specifically, both encoders will be evaluated using the same physiological anchor targets and comparable training configurations. We will report their prior-fitting MAE and RMSE, together with the corresponding optimization settings and training curves.

This comparison will help determine whether the observed performance difference is associated with the KAN parameterization or potentially arises from differences in prior-fitting quality and optimization procedures.

We will also clarify the limitations of the existing ablation results and distinguish the contribution of KAN-based encoding from that of physiological prior supervision.

W3. Limited Validation of Physiological Consistency and Generalization

We agree that classifier-based age decoding alone cannot fully establish whether generated EEG exhibits appropriate age-dependent physiological changes.

To provide more direct evidence, we plan to extend the existing waveform-marker analysis to age-related physiological characteristics.

Specifically, we will examine age-dependent changes in frequency-band power, slow-wave activity, spindle characteristics, K-complex features, mixed-frequency activity, and spectral slope. We will compare the corresponding trends across age conditions and sleep stages, using real EEG as the physiological reference.

Regarding generalization, we acknowledge that the original evaluation is based on 30 subjects from a single dataset and one subject-disjoint split. We will explicitly clarify this limitation and avoid overgeneralizing the current findings. We will also investigate additional subject-disjoint splits to assess the robustness of the observed age–stage consistency, subject to the availability of sufficient computational resources.

AuthorsOpenReview ↗

Questions(1-3) and closing remarks

Q1. Sensitivity to Physiological Anchors, Interaction Coefficients, and Branch Weights

We plan to perform systematic perturbation experiments on the physiological prior parameters.

For the 12 physiological coordinates, we will examine both individual-coordinate and joint perturbations. For the interaction coefficients and six branch weights, we will vary the corresponding parameters while keeping the remaining configurations fixed.

We will evaluate prior-fitting errors, physiological constraint violations, and changes in the learned response functions.

In addition, we will investigate the sensitivity of the prior-distillation loss weights to assess whether the learned physiological representation depends excessively on a particular loss-weight configuration.

These experiments will provide a quantitative assessment of the stability of the handcrafted physiological prior.

Q2. Prior-Fitting Accuracy and Optimization Fairness of Phy_MLP

We thank the reviewer for raising this important question.

We agree that the comparison between Phy_MLP and Phy-DiffKAN should include not only downstream performance but also their ability to approximate the same physiological prior field.

We therefore plan to evaluate both encoders using identical physiological anchor targets and comparable six-branch conditioning configurations.

We will report prior-fitting MAE and RMSE, model parameter counts, optimizer settings, learning rates, training epochs, and hyperparameter configurations.

We will also clarify the optimization procedures used for both variants to ensure that the comparison is appropriately controlled.

This additional analysis will help establish whether the performance difference is attributable to the encoder architecture rather than unequal prior-fitting quality or optimization effort.

Q3. Direct Validation of Age-Dependent Physiological Changes and Robustness Across Subject Splits

We agree that direct physiological analysis and subject-split robustness are important for strengthening the conclusions of this work.

For physiological validation, we plan to fix recorded sex and sleep stage while varying the assigned age. We will extract age-related physiological features from the resulting EEG and compare their trends with those observed in real EEG.

For subject-split robustness, we acknowledge that the current results are based on a single subject-disjoint partition. We will clarify this limitation and investigate additional subject-disjoint evaluations to determine whether the observed improvements are consistent across different subject allocations.

We will report the evaluation protocol and the corresponding performance variations, while ensuring that training and testing subjects remain strictly separated. Closing Remarks

We sincerely thank the reviewer again for the constructive comments. We believe that the proposed sensitivity analyses and controlled encoder comparisons will strengthen the methodological validation of Phy-DiffKAN.

In particular, these additional analyses will help establish the numerical stability of the physiological prior, clarify the prior-fitting and optimization differences between Phy_MLP and Phy-DiffKAN, and provide further evidence regarding age-dependent physiological characteristics.

We will carefully report the experimental protocols, quantitative results, and limitations in the revised manuscript.

Closing Remarks

We sincerely thank the reviewer again for the constructive comments. We believe that the proposed sensitivity analyses and controlled encoder comparisons will strengthen the methodological validation of Phy-DiffKAN.

In particular, these additional analyses will help establish the numerical stability of the physiological prior, clarify the prior-fitting and optimization differences between Phy_MLP and Phy-DiffKAN, and provide further evidence regarding age-dependent physiological characteristics.

We will carefully report the experimental protocols, quantitative results, and limitations in the revised manuscript.

AuthorsOpenReview ↗

Results of question1

W1/Q1: Sensitivity to physiological anchors, coefficients, and branch weights

We thank the reviewer for requesting a quantitative sensitivity analysis. We have now evaluated 46 perturbation settings across 12 physiological anchors, 5 main/interaction coefficient groups, and 6 branch loss weights.

Experimental Setup: All models were retrained under a unified configuration across the same 90 age-sex-sleep-stage conditions (750 epochs, AdamW, lr = 3 × 10⁻⁴, weight decay = 10⁻⁵, gradient clipping = 1.0). The baseline model was trained with 3 random seeds. Three key anchors were repeated with 3 seeds, while the remaining settings used a single seed for descriptive comparison.

Metric Definitions:

  • Prior MAE: Measures how well the model fits the modified targets. It is the mean absolute error between the PhysioKAN output zᵖ and the perturbed prior target τᵖ across 6 branches (b), 90 conditions (c), and 12 features (j): Prior MAE = (1/6) * Σ_b [ (1 / (90*12)) * Σ_{c,j} | zᵖ_{b,c,j} - τᵖ_{b,c,j} | ]
  • Output Shift: Measures how much the model's internal representation changes relative to the original unperturbed baseline z⁰.
    • For Anchors (Relative Shift): We report the mean absolute change normalized by the mean absolute baseline output: Relative Shift = ( |zᵖ_j - z⁰_j|_mean / ( |z⁰_j|_mean + ε ) ) × 100%
    • For Coefficients/Weights (Absolute Shift): We report the unnormalized mean absolute change: Absolute Shift = (1 / (90*12)) * Σ_{c,j} | zᵖ_{b,c,j} - z⁰_{b,c,j} |

1. Prior approximation remains highly accurate across all perturbations. Compared with the unperturbed baseline (Prior MAE = 0.00564 ± 0.00044 across 3 seeds), the perturbed models maintain a comparable fitting error. The data ranges are summarized below:

Perturbation TypeSettingPrior MAEOutput Shift
Physiological Anchors±20%0.00490 – 0.005556.68% – 20.27% (Relative)
Main/Interaction Coeffs0.8×, 1.2×0.00474 – 0.005720.00568 – 0.04355 (Absolute)
Branch Loss Weights0.8×, 1.2×0.00505 – 0.005710.00010 – 0.01328 (Absolute)
Baseline (Unperturbed)N/A0.00564 ± 0.00044N/A

2. Stable fitting error does not mean an insensitive representation. The results demonstrate that while the model accurately fits the new targets, its internal learned coordinates change meaningfully in response to the perturbed parameters:

  • Anchors (±20%): Produce relative output shifts ranging from 6.68% to 20.27%.
  • Coefficients (0.8×/1.2×): Scaling age main effects produces absolute branch-output shifts of 0.04185 – 0.04355 (vs. 0.00877 – 0.00879 for age–stage interactions).
  • Branch Weights (0.8×/1.2×): The stage branch shifts by 0.00010 – 0.00011, while the sex branch shifts by 0.00910 – 0.01328.

Conclusion & Scope: Within the tested ±20% perturbations, PhysioKAN retains comparable prior-fitting accuracy while its branch outputs actively respond to changes in the prior specification and training weights. As stated in Section 3.2, the anchors encode physiological directions on a dimensionless modeling scale; they are not calibrated physical measurements. This analysis confirms the model's local numerical sensitivity and adaptability, distinguishing prior approximation from downstream generation robustness.

Response to W2 (Insufficient isolation of methodological contributions) & W3 (Limited validation of physiological consistency)

We thank the reviewer for these critical points. To explicitly isolate the contribution of physiological knowledge from the network architecture (W2) and to directly validate physiological consistency without relying solely on classifiers (W3), we conducted a capacity-matched No-Prior ablation (identical DiffKAN architecture and conditioning capacity, but trained without the physiological prior).

The results (Tables R1-R3 below) demonstrate that physiological supervision—not merely the network structure—is the fundamental driver of our performance.

1. Isolating Methodological Contributions (W2)

To determine whether improvements arise from physiological knowledge or the KAN parameterization, we evaluated the capacity-matched No-Prior model. As shown in Table R1, removing the physiological prior causes a catastrophic collapse across all metrics. Crucially, the FID skyrockets from 0.53 to 30.88, and the LogPSD Error surges from 0.06 to 48.25. The REM age decodability drops to near random chance (0.511).

This proves that the DiffKAN architecture alone cannot generate coherent data; the physiological prior provides the essential structural guidance.

Table R1: Performance Comparison (Original vs. Capacity-Matched No-Prior)

Evaluation MetricOriginal Phy-DiffKANNo-Prior (Capacity-Matched)
FID ↓0.53 ± 0.0330.88
LogPSD Error ↓0.06 ± 0.0148.25
Five-stage Balanced Acc ↑0.87 ± 0.010.764
REM Age Balanced Acc ↑0.72 ± 0.000.511
Joint Consistency ↑0.5460.344

2. Direct Validation of Physiological Consistency (W3)

You correctly noted that relying on pretrained classifiers is insufficient. To directly validate age-dependent physiological changes, we analyzed the expected directional changes of specific EEG frequency bands. As shown in Table R3, our Original model aligns with all physiological expectations (e.g., decreased Delta, Theta, and Slow-Wave activity, and increased Beta power).

In contrast, the No-Prior baseline fails to capture the correct physiological direction for critical features: it reverses the trend for Sigma Power (0.075 instead of the expected decrease). This directly validates our physiological consistency and proves that the prior is necessary to learn the correct age-stage modulation.

Table R3: Age-Related Physiological Feature Correlations (No-Prior)

Physiological FeatureAge Correlation (ρ)Expected DirectionAgreement
Delta Power-0.398DecreaseYes
Theta Power-0.331DecreaseYes
Alpha Power-0.059DecreaseYes
Sigma Power+0.075DecreaseNo
Beta Power+0.127IncreaseYes
Slow-Wave Activity-0.363DecreaseYes

3. Baseline Sample Quality and Retention

Furthermore, to address the implicit question of why the No-Prior model struggles with conditional consistency, we examined its stage classification performance. As shown in Table R2, the No-Prior model completely fails to generate valid REM samples, achieving an extremely low REM Recall (0.440) and F1-score (0.537).

Because its generated REM samples are largely unidentifiable, its strict sample-retention rate for coherent age-stage pairs is near zero, making it impossible to maintain age-stage consistency.

Table R2: Sleep Stage Classification Performance of No-Prior

Sleep StageRecallF1-score
Wake0.9780.898
N10.6360.625
N20.9110.774
N30.8550.922
REM0.4400.537
Macro Average0.7640.751

(Note: Responses regarding W1/Q1 (Sensitivity to priors) and W2's Phy_MLP comparison have been addressed in our previous responses.)

We hope these new ablations and direct physiological validations fully clarify the contribution of our method.

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

Related papers

Discussion (0)

Sign in to comment.