acceptodds
Under review as a conference paper at ICLR 2027

Cyclostationary Phase Conditioning for Medical Time Series Diffusion

Abstract

Many physiological time series, such as cardiac and brain recordings, exhibit cyclostationarity: their statistics vary periodically with an underlying cycle phase. Corruption from motion, poor contact, and physiological interference obscures morphology needed for diagnosis, making signal restoration essential. Existing diffusion approaches condition on corrupted observations alone and must learn cyclic structure implicitly. We instead propose two inductive biases which encode cyclostationarity: a shift-covariant wavelet representation and dense per-sample phase conditioning inferred from the corrupted input. We further introduce a training-free cyclostationarity index that quantifies phase structure and predicts when phase conditioning will help. Finally, we propose antithetic coupling of reverse trajectories to reduce sampling variance while achieving comparable performance with fivefold fewer network evaluations. Across modalities, our results show that explicitly encoding measurable cyclic structure improves physiological time-series restoration.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
diffusion models, time series, cyclostationarity, physiological signals, medical time series, phase conditioning, inductive bias, stationary wavelet transform, antithetic sampling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_dN14OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces a new conditional diffusion framework, PTSDiff, for multiple different physiological time-series restoration. The paper claims that the physiological time-series data contain wide-sense cyclostationarity property, and leveraging this information for training improves the performance. In particular, the model uses shift-covariant representation to satisfy that property and uses phase information as additional conditioning input. The results show that PTSDiff achieves the best performance in most of the datasets on three different metrics.

Critical strengths

The paper is well-written and easy to follow. The conditional diffusion on shift-covariant wavelet representation model design is clean and novel. The experiments provide a comprehensive evaluation and show superior performance compared with other baselines.

Critical weaknesses

Unclear contribution of phase conditioning: the two main novelties of PTSDiff are “training diffusion on a shift-covariant wavelet representation” and “using dense per-sample phase conditioning inferred from the corrupted input”. However, the phase conditioning contribution is unclear.

  1. Only ECG phase encoder is warm-started toward phase targets. Other three modalities directly learn the phase encoder from the corrupted data y. Therefore, it is unclear if the phase encoder really learns the phase information or simply acts as an additional feature encoder for the denoiser. Could the authors verify what the learned channels represent, e.g., by comparing the predicted phase with detector-derived phase?
  2. The phase-shuffle experiment in Figure 2 is to support the claim that “phase conditioning helps more when cycle structure is stronger”. However, permuting any informative per-sample conditioning would reduce performance, so the drop shows that the model relies on the branch, not that the branch encodes phase or how much it helps. The retrained ablation in Table 17 is a clearer measure of the benefit, and it shows a much smaller effect than the shuffle drop. In addition, the y-axis of Figure 2(b) (the ΔSNR margin over the best baseline) also depends on the strength of each modality’s baseline, and the correlation rests on only four points. In my opinion, plotting the Table 17 benefit on using phase conditioning against the cyclostationarity index in Table 18 would more directly support the claim.
Questions for authors

Clarification requests:

  1. According to section 3.2, “the undecimated transform is a tight frame with A = B = J+1”. However, the provided code uses the unnormalized SWT from pywt, which should give us the frame bounds A = 2, B = 2^J = 16. The frame is therefore not tight and the constant in equation 11 needs to be corrected.

Suggestions for future work:

  1. PTSDiff uses SWT coefficients for diffusion training, and the benefits are clear that it is shift-covariant. However, it is redundant as mentioned in section 3.2. The valid coefficients form only an L-dimensional subspace of the 5L space. However, diffusion model adds and predicts noise on all 5L dimensions, and the noise in the remaining 4L directions is exactly predictable from x_t^F. Can the author explain why this redundancy is not an issue? For example, if the predicted clean coefficients \hat{x}_0^F were projected onto the valid coefficients at every reverse step, i.e. \hat{x}_0^F <- swt(iswt(\hat{x}_0^F)), which would set the 4L components to 0, would it improve or hurt the model? I think the effect of the redundant space on diffusion is worth analyzing.
  2. Inference computation: Table 21 compares the model parameter, NFE, and wall-clock time with DeScoD-ECG and TFCDiff, but only on ECG. The deterministic baselines (DDAE, TCDAE, FCN, EEGIFNet) require a single forward pass (NFE = 1), and the stochastic baselines on the other modalities (SDEMG, EEGDfus) appear to use a single sample rather than MC-K or AV-K averaging. It would be better to also include the NFE and inference time for all baselines, and either apply averaging to the stochastic baselines or report PTSDiff without averaging. It is fine to have larger inference computation for better performance, but I believe this trade-off is worth mentioning.
  3. Training computation: It would be better to include the training time as well to help readers gauge the computational requirements.
3 comments
AuthorsOpenReview ↗

Response to Reviewer dN14 - Part 1

We thank the reviewer for their thoughtful feedback and suggestions. Below, we address each point individually.

Clarification of Physiological Phase Conditioning (Weakness 1)

We agree that, without explicit phase supervision, the learned encoder is not guaranteed to recover interpretable physiological phase.

For ECG, where a reliable analytical peak detector is available, we initialize the encoder with detector-derived phase targets before jointly optimizing it for restoration. We now provide an analysis of phase-estimation errors throughout training Appendix D.15, Figure 12) to assess how closely the learned representation remains aligned with physiological phase. Furthermore, a new matched-capacity ablation (Appendix D.8, Table 20) shows that phase-informed initialization improves restoration compared with training the same encoder from scratch, demonstrating the benefit of incorporating explicit physiological phase information beyond additional model capacity.

For PPG, EMG, and EEG, where no explicit phase supervision is used, we acknowledge that the learned channels may represent more general temporal features rather than physiological phase. We therefore revise the title and manuscript to distinguish phase-informed initialization for ECG from phase-inspired learned temporal conditioning for the other modalities.

Extending explicit phase supervision to these modalities is not straightforward and remains future work, because none of them offers a reliable analytical phase target comparable to the ECG R-peak. PPG has a well-defined cardiac cycle, but beat detectors degrade under the motion artifacts that dominate wrist recordings (Charlton et al., 2022), and borrowing targets from a synchronous ECG would require a second sensor and a pulse arrival delay that itself varies from beat to beat. Surface EMG is the superposition of many motor units firing at different rates, so it has no single cycle to supervise, and its firing structure is recoverable only by decomposing multi-channel recordings (Farina et al., 2004; Holobar & Zazula, 2007), which our single-channel setting does not provide. EEG oscillations are bursty, non-sinusoidal and band-specific, so a phase exists only within a band and only while an oscillation is present (Cole & Voytek, 2017; Donoghue et al., 2020), which rules out a single dense target of the kind in Eq. (7). We have made this clear in the revised version and added to the future work section.

References

  • Charlton et al., “Detecting beats in the photoplethysmogram: benchmarking open-source algorithms,” Physiological Measurement, 2022.
  • Farina, Merletti and Enoka, “The extraction of neural strategies from the surface EMG,” Journal of Applied Physiology, 2004.
  • Holobar and Zazula, “Multichannel blind source separation using convolution kernel compensation,” IEEE Transactions on Signal Processing, 2007.
  • Cole and Voytek, “Brain oscillations and the importance of waveform shape,” Trends in Cognitive Sciences, 2017.
  • Donoghue et al., “Parameterizing neural power spectra into periodic and aperiodic components,” Nature Neuroscience, 2020.

Phase-Shuffling Experiment and Revised Figure 2 (Weakness 2)

We agree on all points. The shuffle measures how much the trained model relies on the branch, not what the branch encodes or how much it helps, and the margin over the best baseline also reflects baseline strength. We have made the change you suggested: Figure 2(b) now plots the gain from phase conditioning in the retrained ablation (full model minus no-phase model, Table 18) against the cyclostationarity proxy ΠAC\Pi_{\mathrm{AC}} (Table 21, modality mean). The old panel is removed, and the text now presents the shuffle only as a measure of reliance. The panel still has only four modalities, so we use it to show the trend that modalities with stronger cycle structure gain more, not as conclusive statistical evidence. Additional supporting evidence for the trend is the synthetic sweep, with 21 controlled conditions (Figure 3, r = 0.85).

Clarification of SWT Frame Bounds and Eq. (11)

Regarding Clarification Request 1, we thank the reviewer for checking this against the code. You are correct: we use the unnormalized SWT, whose frame bounds are A=2A = 2 and B=2J=16B = 2^J = 16, so the frame is not tight and our statement A=B=J+1A = B = J+1 was wrong. We apologize for the oversight and appreciate the comment.

We have corrected Section 3.2 and the text following Eq. 11. The bound keeps its form, but its constant is now A−1/2=2−1/2A^{-1/2} = 2^{-1/2} instead of (J+1)−1/2(J+1)^{-1/2}, as justified by the operator norm of the synthesis operator used in our implementation (Appendix C.2). A new paragraph in Appendix C.2 gives the filters, the synthesis operator, the frame bounds and a short justification of this norm. The implementation and all reported experimental results are unchanged, since the correction only affects the theoretical analysis motivating the auxiliary losses.

AuthorsOpenReview ↗

Response to Reviewer dN14 - Part 2

Redundancy in SWT-Domain Diffusion (Suggested Future Work 1)

We thank the reviewer for this interesting suggestion. We agree that the redundant SWT representation introduces additional dimensions, including noise components outside the valid coefficient subspace, potentially increasing computational cost. However, this does not prevent valid time-domain reconstruction, as the synthesis operator maps predicted coefficients back to the original signal space. Projecting the predicted clean coefficients onto the valid SWT subspace at each reverse step would enforce coefficient consistency, but could also alter the learned reverse dynamics. Whether this improves restoration quality requires empirical investigation, and we consider it an interesting direction for future work.

Meanwhile, we provide additional ablations comparing SWT against alternative representations (Appendix D.12, Table 24), supporting its empirical effectiveness despite the redundancy.

Inference Computation (Suggested Future Work 2)

Thank you for the suggestion. We agree that the comparison should be complete and the trade-off stated. One clarification first: the SDEMG and EEGDfus rows of Table 1 were already averages over ten trajectories, the same protocol as DeScoD-10 and TFCDiff-10, but their labels and our protocol description did not say so. We have renamed them SDEMG-10 and EEGDfus-10 and corrected the text for more clarity. All stochastic baselines are therefore compared under the same averaging as PTSDiff.

Table 25 (previously Table 21, ECG only) now covers every model of Table 1 in all four modalities, with NFE, parameters and wall-clock time per segment on a single NVIDIA RTX 4090 at batch size 32. As you suggested, it also reports PTSDiff without averaging (MC-1). For completeness we add one antithetic pair (AV-2) for PTSDiff and for every diffusion baseline. All values are means over three training seeds.

With one antithetic pair, PTSDiff exceeds every baseline in all four modalities, at a lower wall-clock time than each averaged diffusion baseline. The single-pass models remain far cheaper (0.1 to 0.4 ms against 19 to 278 ms) in terms of latency; the extra cost buys 3.3 to 4.2 dB on PPG, ECG and EMG but only 0.3 dB on EEG. We now state this inference-time to performance trade-off next to Table 25 in Appendix Section D.14.

Training Computation (Suggested Future Work 3)

Thank you for the suggestion. The revision now reports the training cost of PTSDiff and each baseline in Table 6 (Appendix C.2), measured on a single NVIDIA RTX 4090. The training part for our model is reproduced here:

ModalityDatasetParams [M]Segmentss/epochEpochsTotal [h]
ECGPTB-XL3.4417,441405005.8
PPGPPG-DaLiA3.4421,284353002.9
EEGEEGdenoiseNet9.8829,920574206.7
EMGNinaPro DB23.44159,96014116023.5

A single training run takes between 2.9 h (PPG) and 23.5 h (EMG) on the GPU. Training all four models once takes about 39 GPU-hours in total (the sum of the last column). EMG dominates because of its longer windows (L = 10,000) and a training set five to nine times larger than the others. Each total is the median duration of three timed epochs, measured after one warm-up epoch, multiplied by the configured number of epochs. Totals exclude per-epoch validation; the ECG total includes the phase-encoder warm-start (0.3 h).

Thank you for the authors' response. Most of my concerns has been resolved, including the SWT frame bound, inference computation, training computation, and the phase-shuffling experiment.

Regarding the "Redundancy in SWT-Domain Diffusion", the authors provide a new ablation on using different transformed representation to show the advantage of SWT-domain diffusion, and the results are clear. The authors list the redundancy analysis as future work.

Regarding the "Physiological Phase Conditioning", the new matched-capacity ablation on ECG supports the benefit of phase-informed initialization. For the other three modalities, the authors explain the difficulty of extending explicit phase supervision and revise the claim to "phase-inspired learned temporal conditioning". Although this makes the paper's claim weaker, it is now more honest and better supported by the evidence.

In conclusion, the authors' response resolves most of my concerns, with the remaining questions either deferred to future work or addressed by narrowing the claims. I am happy to remain my score of 3 (Weak acceptance).

Reviewer_Pq6BOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PTSDiff, a conditional diffusion model for restoring corrupted physiological time series. Its components are: (a) stationary wavelet transform (SWT) diffusion preserving temporal alignment across scales, (b) dense learned conditioning intended to encode cycle phase, event location, and local cycle rate, (c) a cyclostationarity index measuring the fraction of variance explained by phase, proposed as an indicator of when phase conditioning helps, and (d) antithetic reverse sampling, pairing opposite-noise trajectories to reduce reconstruction variance, plus a global context encoder and time/morphology losses. Validation spans PPG, ECG, EMG, and EEG with modality-specific baselines, three seeds for the main comparison, ablations, synthetic experiments, downstream evaluations, and held-out datasets. Gains are substantial on cardiac modalities and EMG; EEG is mixed. Antithetic sampling improves competing diffusion restorers without retraining.

Critical strengths

(a) Substantial restoration gains supported by downstream and cross-dataset evaluation. ECG Δ\DeltaSNR improves from 13.8513.85 dB (DeScoD) to 15.8715.87 dB in Table 1, while downstream MIT-BIH F1 improves from the strongest baseline's 0.5430.543 to 0.6300.630, with improvements persisting on held-out cardiac datasets. This connects waveform restoration to downstream behavior beyond a single dataset.

(b) Convincing practical evaluation of antithetic sampling. Matched-NFE analysis, direct correlation measurements, wall-clock results, and sampler swaps on competing methods; Table 21 shows an antithetic pair can match or exceed ten independent trajectories at substantially lower cost. The contribution is primarily validation and integration here, with earlier use in physiological diffusion restoration explicitly acknowledged.

Critical weaknesses

(a) The experiments do not isolate physiological phase conditioning from generic additional representation capacity. For PPG, EMG, and EEG the phase encoder is trained solely through restoration loss; no reported constraint ensures monotone phase, positive cycle rate, valid sine/cosine relationships, or event localization. ECG receives phase supervision during initialization, but subsequent training can change the representation. Removing the branch shows additional conditioning is useful; shuffling it shows the model uses it; neither shows the benefit comes specifically from physiological phase. A matched-capacity generic conditioning encoder and direct phase-quality measurements are important controls. This is central to the recommendation because phase-specific inductive bias is the main conceptual contribution.

(b) The mathematical account of wavelet and symmetry benefits overstates what is established. The unsupported frame constant propagates into Eq. (11), and SWT equivariance is used to motivate stronger claims about a strided architecture. These concern the stated foundation of one of the two principal inductive biases. The empirical gains could survive corrected mathematics, but the current submission does not establish whether they arise from shift symmetry, multiscale filtering, redundancy, or other architectural changes.

Questions for authors

(a) What enforces the interpretation of the five learned channels? Are phase, sine/cosine consistency, event localization, and rate constrained in the implementation? What phase or event-estimation accuracy remains after joint restoration training?

(b) What are the exact SWT analysis and synthesis operators? Please specify filter normalization, boundary handling, and scale weighting, and establish applicable frame bounds. How are arbitrary-shift equivariance claims reconciled with stride-2 downsampling?

(c) Which index is plotted in Figure 3: theoretical phase-explained variance, a detector-based estimate, or the autocorrelation proxy? How are phase, normalization, and cycle-length jitter handled? This matters because phase alignment should remove pure timing variation from a fixed cyclic template under suitable assumptions.

(d) Does the phase-conditioned model outperform an equally sized generic conditioning branch under matched training and supervision? Together with phase-quality measurements, this would determine whether the central phase-specific explanation is supported.

Should the authors adequately address all concerns and questions, I am willing to raise my score.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Pq6B - Part 1

We thank the reviewer for their thoughtful feedback and suggestions. Below, we address each point individually.

Clarification of Physiological Phase Conditioning and Encoder Capacity (Weakness a, Questions a & d)

We agree that the learned encoder is not guaranteed to recover interpretable physiological phase without explicit supervision, and that the phase-shuffling experiment establishes reliance on the conditioning branch rather than its physiological interpretation. The five channels are supervised to represent event localization, cycle phase, sine/cosine embeddings, and local cycle rate only during ECG warm-starting; no explicit constraints enforce these interpretations or their consistency during subsequent joint training.

ECG is the modality where reliable analytical phase targets are available, and we use it as a controlled test-bed to isolate the effect of the phase contribution. To study the contribution of phase-informed initialization, we introduced a matched-capacity ablation comparing identical architectures trained with the same restoration objective, with and without phase-informed warm-starting (Appendix D.8, Table 20). The phase-informed initialization improves restoration by 0.49 dB demonstrating a benefit beyond just additional encoder capacity. As control, we also report phase-estimation errors throughout joint training (Appendix D.15, Figure 12). The phase error remains consistently low (approximately 0.02–0.03 cycles), even after the encoder is unfrozen and optimized solely through the restoration objective, and remains substantially below both the analytical detector on corrupted inputs (0.10 cycles) and chance level (0.25 cycles). This confirms that the learned representation retains meaningful physiological phase information throughout training.

For PPG, EMG, and EEG, where reliable analytical phase supervision is not readily available, we acknowledge that the learned channels may represent more general temporal features rather than physiological phase. Extending explicit phase supervision to these modalities is not straightforward and remains future work, because none of them offers a reliable analytical phase target comparable to the ECG R-peak. PPG has a well-defined cardiac cycle, but beat detectors degrade under the motion artifacts that dominate wrist recordings (Charlton et al., 2022), and borrowing targets from a synchronous ECG would require a second sensor and a pulse arrival delay that itself varies from beat to beat. Surface EMG is the superposition of many motor units firing at different rates, so it has no single cycle to supervise, and its firing structure is recoverable only by decomposing multi-channel recordings (Farina et al., 2004; Holobar & Zazula, 2007), which our single-channel setting does not provide. EEG oscillations are bursty, non-sinusoidal and band-specific, so a phase exists only within a band and only while an oscillation is present (Cole & Voytek, 2017; Donoghue et al., 2020), which rules out a single dense target of the kind in Eq. (7). We have made this clear in the revised version and added to the future work section.

Overall, the updated three-seed, leave-one-out ablations (Table 3) and additional component analyses (Appendix Tables 18–19) establish the benefit of conditioning, but not necessarily its phase-specific interpretation. We have revised the manuscript accordingly, clarifying the distinction between phase-informed initialization on ECG and unsupervised learned conditioning in other modalities.

References

  • Charlton et al., “Detecting beats in the photoplethysmogram: benchmarking open-source algorithms,” Physiological Measurement, 2022.
  • Farina, Merletti and Enoka, “The extraction of neural strategies from the surface EMG,” Journal of Applied Physiology, 2004.
  • Holobar and Zazula, “Multichannel blind source separation using convolution kernel compensation,” IEEE Transactions on Signal Processing, 2007.
  • Cole and Voytek, “Brain oscillations and the importance of waveform shape,” Trends in Cognitive Sciences, 2017.
  • Donoghue et al., “Parameterizing neural power spectra into periodic and aperiodic components,” Nature Neuroscience, 2020.

Clarification of SWT Mathematics and Performance Attribution (Weakness b)

We thank the reviewer for highlighting these concerns. We have clarified the SWT frame bounds ((A=2), (B=16)), explicitly defined the synthesis operator, and clarified the justification for Eq. (11) accordingly (Appendix C.2). We also clarify that shiftability is a property of the SWT representation, rather than the full architecture, whose strided layers do not preserve equivariance under arbitrary shifts. Additionally, we include new ablations comparing SWT against Fourier, DCT, and learned CNN representations (Appendix D.12, Table 24) to assess the effect of the transform choice. We now present shiftability as a motivating property without attributing the observed gains exclusively to it.

AuthorsOpenReview ↗

Response to Reviewer Pq6B - Part 2

Clarification of SWT Implementation, Frame Bounds, and Shiftability (Question b)

Analysis uses the orthonormal PyWavelets filters sym4 (ECG, PPG, EMG) and db4 (EEG), length 8, normalized to ∑kh[k]=2\sum_k h[k] = \sqrt{2} and ∥h∥2=1\lVert h\rVert_2 = 1, with quadrature-mirror high-pass g[k]=(−1)k+1h[7−k]g[k] = (-1)^{k+1} h[7-k]. At level jj, the filters are dilated by 2j−12^{j-1} (à trous) and applied by circular convolution without decimation (Eq. 3); synthesis inverts level by level, aj−1=12(Hj⊤aj+Gj⊤dj)a_{j-1} = \frac{1}{2}(H_j^\top a_j + G_j^\top d_j), where HjH_j and GjG_j denote circular convolution with the dilated filters. Since W\mathcal{W} is non-square, we replace the previous notation W−1\mathcal{W}^{-1} with the synthesis operator S\mathcal{S}, an exact left inverse satisfying SW=I\mathcal{S}\mathcal{W}=I. Boundaries are periodic; the window lengths used in our main experiments (512, 3600, 10000) are multiples of 2J=162^J=16, so no padding occurs. No per-band weighting is applied: the diffusion loss weights all bands equally, and L_rec and L_morph are computed in the time domain after S\mathcal S.

Frame bounds: Without rescaling, band jj carries a gain of 2j/22^{j/2} relative to the Parseval-normalized transform, so A=2A=2 and B=2J=16B=2^J=16, both attained; we verified both numerically for sym4 and db4. S\mathcal{S} has operator norm A−1/2=2−1/2A^{-1/2}=2^{-1/2}, so frame-domain errors are not amplified in the time domain.

We agree with the reviewer that arbitrary-shift equivariance cannot hold for the full network. Eq. (4) establishes shiftability of the representation: W\mathcal{W} and S\mathcal{S} commute with every integer cyclic shift under periodic boundary conditions. The phase encoder is shift-equivariant up to boundary effects, while the U-Net's three stride-2 downsampling stages restrict equivariance to shifts that are multiples of 88, again up to boundary effects. Thus, the full architecture is not exactly shift-equivariant, but the SWT representation introduces no additional shift dependence, unlike a decimated DWT. Appendix C.2 now documents these details and clarifications.


Clarification of the Synthetic Jitter Experiment (Question c)

Figure 3 plots the bias-corrected, phase-aligned estimate Π^\hat{\Pi} of App. B.3, computed using ground-truth event times rather than a detector, and not the theoretical phase-explained variance or the autocorrelation proxy. In the synthetic sweep, the detector is replaced by the generator's ground-truth event times (template peaks), and phase follows Eq. (6) between them. The generator z-scores each window (each component and the mixture), yielding Π^≈c\hat{\Pi} \approx c at zero jitter; Π^\hat{\Pi} additionally normalizes each cycle by amplitude (Appendix B.3). Cycle lengths have mean PP samples and coefficient of variation σjit\sigma_{\mathrm{jit}}. Every event carries the same template, lasting 0.85P0.85P samples (85% of the mean cycle length) and never stretched to the cycle.

We agree that phase alignment would remove pure timing variation if the template scaled with the cycle. Ours deliberately does not, as in ECG, where QRS duration barely depends on the RR interval. Under Eq. (6), the template therefore occupies a different phase interval in each cycle, and cycles shorter than 0.85P0.85P make consecutive templates overlap. Jitter thus lowers the phase-explained variance itself, and Π^\hat{\Pi} tracks it. To clarify this in the manuscript, Section 5, the Figure 3 caption and its x-axis now name Π^\hat{\Pi} as the plotted index, and Appendix D.10 now specifies the ground-truth events, normalization, the fixed-duration template and why Π^\hat{\Pi} falls with jitter.

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

The paper presents a new diffusion-based algorithm for restoring time series, which introduces three new ideas with respect to existing methods: i) the algorithm uses the stationary wavelet transform, which is shiftable; ii) it proposes a conditioning technique based on the phase (the position of the window relative to the cycle), iii) it leverages antithetic sampling to reduce the variance of the averaged posterior samples. The paper also presents some theory illustrating why conditioning on the phase reduces the variance of the cyclostationary signals, and how antithetic sampling can help to reduce the variance of the averaged posterior samples. Experiments on real EEG, ECG, and PPG time-series data show better results than competing diffusion-based methods.

Critical strengths
  • Experiments show a significant performance gap with competitors across various applications
  • The paper is well written and easy to follow, and the theory is well explained and motivates the design choices.
Critical weaknesses
  • It is unclear to me what the motivation is behind using the wavelet transform. The transform is equivalent to a first 'depthwise' convolutional layer in the network with carefully selected filters. There is nothing aside from shift-equivariance, which also holds in CNN layers and is useful for almost any type of 1D, 2D, or 3D signal, that makes this choice particularly well-suited for cyclostationary signals.

    • I believe the idea of using wavelets at the input of a network has resurfaced multiple times in the last decade, but it has never really stuck. I have a hard time imagining how this paper can change this trend. Some examples: "Multilevel Wavelet Decomposition Network for Interpretable Time Series Analysis" (KDD 2018), "Enhancing Foundation Models for Time Series Forecasting via Wavelet-based Tokenization (WaveToken)" (ICLR 2025), "Deep Scattering Spectrum with Deep Neural Networks" (ICASSP 2014), "Wavelet Flow: Fast Training of High Resolution Normalizing Flows" (NeurIPS 2020), "Wavelet Score-Based Generative Modeling" (NeurIPS 2022).
    • I have a hard time understanding the benefits from Table 3: it's quite surprising that most performances fluctuate around .2 dB difference, except when all the modules are added, which jumps to .5 dB or more.
    • the notation W−1\mathcal{W}^{-1} is a slight abuse of notation, the linear operator is not square so there is no well-defined inverse (its more the adjoint you use later on).
    • "is exactly shift-covariant, or equivalently translation-equivariant" - there are multiple confusing points here: the standard discrete wavelet transform is shift-equivariant and translation-equivariant in the strict mathematical definition, ie for the group of transformations Tgg∈G\\{T_g\\}_{g\in \mathcal{G}}, there exist a valid linear group action T~g\tilde{T}_g and TgT_g such that T~gW=WTg\tilde{T}_g \mathcal{W} = \mathcal{W} T_g. This is actually true for any invertible transformation (thus including the discrete wavelet transform). What is meant by the paper (if I understood well) is the "shiftability property", ie that T~g\tilde{T}_g is also a shift operation (which is not the case for the discrete wavelet transform). The other confusion is mixing 'shifts' with 'translations' (the latter typically takes into account sub-pixel/sub-sample shifts).
  • The paper only shows posterior averaging results, and it seems to be the sole and main goal of the algorithm (also motivating the antaithetic sampling strategy). If this is the case, I struggle to understand why a diffusion model is necessary? A direct regressor provides the posterior mean in a single NFE vs 100s of NFEs for the posterior sampler case.

Questions for authors
  • It is hard to understand whether the better performance of PTSDIFF is due to a better architecture/optimizer/data preprocessing/etc than competitors, or if it is due to the claimed phase conditioning and wavelet transform. Looking at Table 3, it seems that most of the gap is not due to the main contributions of the paper (which fluctuate around .5 dB, vs the 2 dB gains over baselines), but rather an overall better training/model architecture strategy?
4 comments
AuthorsOpenReview ↗

Response to Reviewer UnCg - Part 1

We thank the reviewer for their thoughtful feedback and suggestions. Below, we address each point individually.

Motivation behind the wavelet transform and comparison to CNN layers

We use the stationary wavelet transform (SWT) to obtain a temporally aligned, multiscale representation that captures both low-frequency trends and localized high-frequency features, which is particularly suitable for physiological signals with periodic structure and transient morphology (Addison, 2005).

While wavelets can be implemented using convolutions, the SWT provides a fixed, analytically invertible representation without requiring the analysis filters or reconstruction operator to be learned. Unlike learned CNN representations, it therefore introduces a structured inductive bias without additional learnable parameters. Its effectiveness for physiological times series (like ECG) denoising and morphological preservation has also been shown in prior work (Kumar et al., 2021; Ádám et al., 2025). We have clarified this motivation in the manuscript (Section 3.2) and expanded the related work (Section 2) to include the wavelet-based approaches highlighted by the reviewer.

We are aware of other alternatives. To validate our choice, we have added an additional ablation below (and updated Appendix D.12, Table 24) comparing the SWT against the discrete cosine transform (DCT), Fourier transform (FT), and learned convolutional representations (CNN). Across modalities, SWT outperforms the learned CNN representation on ECG, EMG, and EEG, while achieving comparable performance on PPG. This suggests that its fixed, invertible multiscale representation provides a useful inductive bias beyond what the learned convolutional filters capture.

RepresentationPPGECGEMGEEG
SWT12.99 ± 0.0816.00 ± 0.2122.55 ± 0.0515.50 ± 0.06
DCT11.70 ± 0.0815.75 ± 0.1922.02 ± 0.0513.56 ± 0.04
FT11.41 ± 0.0715.43 ± 0.1921.36 ± 0.0412.71 ± 0.04
CNN13.01 ± 0.0815.37 ± 0.2021.13 ± 0.0415.31 ± 0.06

References

  • Addison, P. S. (2005). Wavelet transforms and the ECG: a review. Physiological Measurement, 26(5), R155–R199.
  • Kumar, A., Tomar, H., Mehla, V. K., Komaragiri, R., & Kumar, M. (2021). Stationary wavelet transform based ECG signal denoising method. ISA Transactions, 114, 251–262.
  • Ádám, N., Val’ko, D., Balogh, Z., Madoš, B., & Hurtuk, J. (2025). Comparative evaluation of filtration techniques for ECG signal denoising with emphasis on stationary wavelet transform. Scientific Reports, 15(1), 42514.
AuthorsOpenReview ↗

Response to Reviewer UnCg - Part 2

Clarifications on Component Contributions and Performance Improvements

We thank the reviewer for raising these concerns, both regarding the interpretation of Table 3 in the Critical Weaknesses and the attribution of performance gains in the Questions for Authors.

The original Table 3 compared different combinations of components. We agree that individual contributions were difficult to interpret. We have therefore restructured the table as a systematic leave-one-out ablation, where each row removes exactly one component from the full model. Furthermore, we now also report results over three training seeds instead of using a single seed, and include the paired performance difference relative to the full model:

RowRemoved componentΔSNR [dB]Δ vs. PTSDiff [dB]
PTSDiffnone15.78 ± 0.12—
i.a)Shift-covariant frame15.42 ± 0.09−0.36 ± 0.06
i.b)Phase conditioning15.13 ± 0.05−0.65 ± 0.17
i.c)Context encoder15.30 ± 0.13−0.47 ± 0.16
i.d)Auxiliary losses15.49 ± 0.05−0.28 ± 0.09

The three-seed results also clarify the pattern noted by the reviewer. In the original single-seed results, removing individual components reduced performance by 0.48–0.72 dB; whereas the three-seed results show a more gradual range of 0.28–0.65 dB, supporting the consistent contribution of each component.

To further isolate the contribution of phase conditioning, we introduce an additional ablation (Appendix Table 19), evaluating its effect across all combinations of context encoding and auxiliary losses. Removing phase consistently reduces ΔSNR by 0.23–0.64 dB, confirming that its benefit is not restricted to the full model. Together with the SWT ablation (−0.36 dB), these results provide evidence that both proposed inductive biases contribute meaningfully to performance.

BlockContext encoderAuxiliary lossesΔSNR with phase [dB]ΔSNR without phase [dB]Δ from removing phase [dB]
i✓✓16.0015.36−0.64
ii✓✗15.5214.98−0.54
iii✗✓15.2815.05−0.23
iv✗✗15.2614.75−0.51

Regarding the comparison with existing methods, all baselines were retrained using the same data splits, preprocessing, and evaluation protocol as PTSDiff, while preserving their original architectures, objectives, and recommended training settings wherever possible (Appendix C.3). While the proposed inductive biases explain part of the approximately 2 dB improvement, we acknowledge that other architectural and training choices may also contribute to the remaining gap, and have clarified this in the revised manuscript.


SWT Inverse Notation

We thank the reviewer for pointing this out. We have replaced W−1\mathcal W^{-1} with the synthesis operator S\mathcal S, explicitly defined as a left inverse satisfying SW=I\mathcal S\mathcal W=I, which guarantees exact reconstruction despite the redundancy of the SWT representation. We have updated the notation throughout the manuscript, with further details provided in Appendix C.2.

Clarification of Shiftability versus Equivariance

Indeed, our intended property is shiftability: an integer-sample shift of the input produces the same shift in the SWT coefficients, unlike the decimated DWT. We have revised the manuscript to make this distinction clearer.

Motivation for Diffusion Models over Direct Regression

Diffusion was a deliberate choice (Section 2). A regressor learns a single deterministic map, whereas conditional diffusion learns the conditional distribution, and its averaged estimate is the state of the art for this task: DeScoD-ECG, SDEMG and EEGDfus each report improvements over deterministic restorers. Our retrained baselines reproduce this on ECG and EMG (Table 1), which is why we build our inductive biases into a diffusion model.

We also compare against single-pass regressors in every modality (TCDAE, DDAE, FCN, EEGIFNet; Table 1). PTSDiff-AV-10 improves ΔSNR over them by 3.4 dB on ECG, 4.3 dB on PPG, 3.4 dB on EMG, and achieves comparable performance on EEG. The ECG margin grows with corruption, from 3.2 to 4.4 dB (Appendix D.2), and persists on downstream tasks (Table 2). We acknowledge that diffusion is computationally more expensive than single-pass regression, but our results demonstrate a substantial improvement in restoration quality. Antithetic sampling reduces the inference cost compared with standard diffusion averaging: a single antithetic pair achieves comparable or better performance than ten independent trajectories with five times fewer network evaluations (Appendix D.14, Table 25). Importantly, averaging is an inference-time choice, not a requirement of our approach, and the model remains a conditional sampler. It allows us to use the averaged estimate when minimizing reconstruction error or an individual sample when a plausible realization of the clean signal is desired.

Many thanks for taking the time to provide an answer to my review.

The ablations are more clear now.

I still have two concerns:

  • did you choose the CNN to match the filter size of the SWT, and to be depthwise? I would expect in that case that the additional degrees of freedom of the CNN filters should outperform the fixed SWT?
  • my point regarding the regressor might not have been clear enough: I think the comparison should be with the same network architecture, and training scheme, but just trained to regress directly the output instead of via a series of diffusion steps. Currently the comparison is with other models, which confounds factors that are orthogonal to being a direct regressor (choice of architecture, dataset, training hyperparams, etc)
AuthorsOpenReview ↗

Response to Follow-Up by Reviewer UnCg

Thank you for the follow-up!

We ran both controls as requested and added them to the revised manuscript (Appendix D.12 and Table 24; Appendix D.14 and Table 25). All new changes are highlighted in green in the uploaded manuscript.

Learned layer matched to the SWT

The CNN in Table 24 used kernel size 8 in every band and stride 2. As suggested, we conducted a more controlled comparison to assess the effect of fixed versus learnable filters, training two alternatives that are both depthwise, undecimated (stride 1), and fully learnable:

  • Dense depthwise convolution: for each input signal (xtx_t, yy and the five channels of Φ^\hat{\Phi}), one filter per band with kernel 106, the length of the SWT’s longest equivalent filter, plus one shared synthesis filter per band (4,240 weights), initialized at the SWT.
  • Learnable SWT: the SWT’s own four-level cascade (kernel 8, dilations 1, 2, 4, 8) with free analysis filters for each input signal and free synthesis filters (512 weights), initialized at the wavelet or at random.

As for the learned convolution in Table 24, the layer is the first and last layer of the denoiser and noise is added in the time domain, so the comparison with PTSDiff also includes this difference. All other settings are those of PTSDiff. We show below the ΔSNR [dB] with one antithetic pair, with mean ± std over three seeds.

First and last layerPPGECGEMGEEG
Fixed SWT frame (PTSDiff)13.19 ± 0.1015.78 ± 0.1222.67 ± 0.1915.36 ± 0.07
Dense kernel 106, SWT init13.15 ± 0.0715.59 ± 0.2122.65 ± 0.0515.36 ± 0.06
Learnable SWT, wavelet init13.00 ± 0.0115.72 ± 0.1322.55 ± 0.0515.32 ± 0.12
Learnable SWT, random init13.28 ± 0.2115.11 ± 0.1122.15 ± 0.0414.90 ± 0.08

The fixed SWT matches or outperforms both learned alternatives across all four modalities, despite requiring no learnable filter parameters. With wavelet initialization, both remain within 0.2 dB of PTSDiff, while random initialization of the learnable SWT reduces performance on ECG, EMG and EEG. Since all variants preserve shiftability, the results suggest that the wavelet filters provide a useful inductive bias beyond shiftability alone. We have revised Section 5 and Appendix D.12 to reflect these findings without claiming that the SWT is an optimal transform.


Same network trained as a direct regressor

We trained the PTSDiff network itself to map yy to x0x_0 in a single pass, with the same U-Net, frame, phase and context encoders, data, optimizer, schedule, epochs, seeds and checkpoint selection. We removed only what requires a diffusion step: the noisy input xtx_t and the step embedding (at most 1% of the parameters). The noise-prediction loss becomes an L2 loss on the clean frame coefficients, and the auxiliary losses are kept (with αˉt=1\bar{\alpha}_t = 1). For comparison, MC-1 denotes a single diffusion trajectory without averaging, while AV-2 averages one antithetic pair. The table reports ΔSNR [dB] as mean ± standard deviation over three seeds:

ModelNFEPPGECGEMGEEG
Best existing single-pass baseline (Table 1)18.97 ± 0.1412.49 ± 0.6519.28 ± 0.2015.05 ± 0.04
PTSDiff network as regressor112.10 ± 0.1415.15 ± 0.0621.95 ± 0.1114.18 ± 0.20
PTSDiff, MC-15011.26 ± 0.1713.52 ± 0.1520.44 ± 0.3014.64 ± 0.05
PTSDiff, AV-210013.19 ± 0.1015.78 ± 0.1222.67 ± 0.1915.36 ± 0.07

The matched regressor already accounts for 2.7–3.1 dB of the 3.3–4.2 dB advantage over existing deterministic baselines on PPG, ECG and EMG, confirming that architectural and training differences contribute substantially to the original margin. Nevertheless, PTSDiff with one antithetic pair consistently outperforms its matched regressor across all four modalities by 0.6–1.2 dB, at 100 versus 1 NFE. We have updated Appendix D.14 and the limitations to reflect this distinction and the associated computational trade-off.


Thank you again for your constructive suggestions, which have strengthened our analysis. We hope the new matched controls address the remaining concerns and support a more favorable overall assessment.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Response

Dear reviewers,

We thank you all for your thoughtful feedback and constructive suggestions. We have carefully addressed the concerns and substantially revised the manuscript, with changes highlighted in blue. We believe that quality and clarity have improved.

The main revisions are:

  1. Phase conditioning: We add experiments assessing the contribution of phase-informed initialization (Table 20) and the physiological interpretation of learned conditioning (Figure 12). We clarify its limitations across modalities.
  2. Wavelet representation: We introduce comparisons with alternative transforms (Table 24) and provide further mathematical and implementation details on SWT.
  3. Component contributions: We extend the ablations across multiple seeds, provide an additional analysis of component interactions, and revise Figure 2 and Table 3 to better assess the benefits of conditioning. We also added supplementary Tables 19, 20, and 24 on component analysis ablations.
  4. Computational efficiency: We extend the inference-cost comparisons to all modalities and baselines, including single-sample and antithetic sampling (Table 25), and provide training-time measurements in supplementary Table 6.
  5. Additional clarifications: We provide further details on the synthetic experiments and clarify the interpretation and limitations of our findings.

You can find further details in our point-by-point responses. All table, figure, and appendix references refer to the revised manuscript in blue.

We look forward to any additional feedback.

Dear Reviewers,

Now that the authors have responded, please review what they have said and give consideration to whether they have adequately addressed your concerns.

-The AC

AuthorsOpenReview ↗

Kind follow-up on answer to reviewers

Dear Reviewers,

Thank you again for your thoughtful feedback. We would like to kindly follow up on our detailed responses and revised manuscript, in which we have carefully worked through your concerns. This is just a reminder, there is no new information or data with respect to our first replies.

In particular, we have:

  1. Added matched-capacity experiments and phase-accuracy analyses, and clarified the interpretation and limitations across modalities (Table 20, Figure 12).
  2. Corrected and clarified the mathematical foundations and added comparisons against alternative representations (Table 24).
  3. Expanded the ablations across multiple seeds and investigated component interactions (Table 3, Tables 18–19).
  4. Extended inference and training cost comparisons across all modalities and baselines (Tables 6 and 25).
  5. Revised the cyclostationarity analyses, synthetic experiments, and interpretation of the results.

All changes are highlighted in blue in the revised manuscript, with further details in our point-by-point responses. We would be very grateful if you could revisit our responses and share any remaining concerns or questions. We would be happy to address them.

Thank you again for your time and constructive suggestions.

Best regards, The Authors

Dear Reviewers,

Now that the authors have responded, please review what they have said and give consideration to whether they have adequately addressed your concerns.

-The AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.