acceptodds
Under review as a conference paper at ICLR 2027

Context-Adaptive Flow Matching for ICU Time-Series Generation

Abstract

Synthetic ICU time-series data can support clinical modeling, but data that help train a forecaster may still distort temporal dependence, spectral structure, or physiological relationships. We propose OU-WSP FM, a flow-matching framework for conditional ICU data generation with a structured endpoint prior. Its Ornstein–Uhlenbeck-inspired autoregressive component models local mean-reverting dynamics, while its Woodbury-spectral component extrapolates the remaining context residual using a low-rank Fourier basis. Together, these components incorporate context-dependent temporal structure into the source distribution without dense covariance inversion. Experiments on MIMIC-III and MIMIC-IV evaluate forecasting utility, statistical fidelity, and blood-pressure ordering, with ablations of spectral strength, channel sharing, and network architecture. The GRU and TCN experiments report composite forecasting errors \(22.0%\)–\(27.2%\) lower than Gaussian FM. To distinguish the effects of endpoint design and calibration, we conduct detailed ablation experiments and calibration analyses under context-only imputation. These findings clarify the distinct roles of structured endpoints, calibration, and clinical safeguards in ICU time-series generation.

Primary area:
generative models
Keywords:
ICU time-series, Synthetic data generation, Flow matching, Ornstein-Uhlenbeck process, Clinical constraints, Downstream utility

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_XnMAOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes OU-WSP FM, a conditional flow-matching framework for ICU time-series generation. The main idea is to replace a standard isotropic Gaussian source with a structured endpoint composed of an OU-inspired autoregressive anchor and a context-adaptive Fourier residual. The model is evaluated on MIMIC-III and MIMIC-IV using downstream forecasting utility, statistical fidelity, and blood-pressure consistency. The paper reports clear improvements over Gaussian flow matching and several generative baselines, particularly in downstream forecasting and spectral fidelity.

Critical strengths

The strongest aspect of the paper is the structured source design. Instead of relying on an uninformative Gaussian endpoint, the method injects temporal information from the observed context before the learned flow transformation begins, which is well motivated for ICU signals. The empirical results are also encouraging: OU-WSP substantially improves over Gaussian FM in downstream forecasting, while the fidelity analysis shows improvements in autocorrelation and spectral metrics. I also appreciate that the authors do not claim that blood-pressure consistency emerges automatically from the generator, and instead explicitly distinguish raw generation from deterministic clinical repair.

Critical weaknesses

The largest weakness is the mismatch between the strength of the main empirical claims and the reliability of the underlying evaluation pipeline. The appendix indicates that the original preprocessing may use future information within the 48-hour window, while the corrected context-only protocol is evaluated only through a narrower source/calibration analysis and does not reproduce the key GRU/TCN downstream results. In addition, some reported fidelity comparisons are affected by inconsistent post-processing: for example, the Gaussian fidelity rows correspond to BP-filtered samples, while post-processing is not uniformly matched across methods. These issues do not invalidate the proposed modeling idea, but they weaken the evidence that the reported gains can be attributed cleanly to OU-WSP under a consistent conditional-generation setting.

Questions for authors

The main question I would like the authors to clarify is whether the headline downstream forecasting improvements remain when the entire preprocessing and generation pipeline is rerun under the strictly context-only protocol described in Sec. 4.6 and Appendix B. At present, the additional analysis helps diagnose the source and calibration behavior, but it does not directly validate the primary forecasting result under the corrected information boundary. I would also like clarification on how the authors intend Table 2 to be interpreted given that post-processing is not uniformly matched across methods, since this affects whether the table should be viewed as a direct fidelity ranking or only as a pipeline-level comparison.

2 comments
AuthorsOpenReview ↗

Response to Reviewer XnMA: Context-Only GRU/TCN Results and Matched Post-Processing

We appreciate the reviewer’s careful reading and the chance to clarify both points. We completed the context-only downstream evaluation, added a three-seed spectral analysis, and a released-TSFlow comparison.

Q1. Downstream evaluation under the context-only protocol

C1 is the completed context-only evaluation, not a narrow source/calibration diagnostic: it covers GRU and TCN on both cohorts (four combinations), all sources, and observed-target scoring. There is no test/validation leakage: generator fits and synthetic training contexts use training-split windows only; validation and test contexts are never used to fit generators or construct synthetic training data. The appendix's 48-hour concern is within-window lookahead in the original imputation, not cross-split leakage; the context-only rebuild removes it. Each synthetic source supplies 6,000 pairs; three repeats are downstream initializations for a fixed generated set. OU-WSP achieves the lowest observed Total among synthetic sources on MIMIC-III (2.1952 GRU, 2.2128 TCN, versus 2.2153/2.2188 for Gaussian FM); across the four combinations it is 0.01%–0.91% below Gaussian FM and 1.54%–2.34% above Real on MIMIC-III. The 22.0%–27.2% comparison used a different protocol; retained as history.

Q2. Fidelity under matched post-processing

R2 reports one matched protocol: 256 held-out contexts × 8 samples, no BP repair or filtering, with output calibration stated per row. OU-WSP has the lowest mean log-PSD discrepancy on both datasets (0.1875 vs 0.2447 on MIMIC-III; 0.1539 vs 0.2119 on MIMIC-IV). Same-backbone rows are three-seed means; TSFlow is one sampling realization.

Spectral evidence and residual increment

R1 reports log-PSD by seed. OU-WSP is lower than Gaussian FM in all six combinations, with mean reductions of 23.4% and 27.4%. Relative to the OU-only control, the reductions are 8.1% and 2.4% (MIMIC-III 3/3 seeds, MIMIC-IV 2/3).

Comparison with adapted TSFlow

We adapted released TSFlow to 24-hour ICU contexts and the 51,200-update recipe. Under C1, OU-WSP is 3.28%–4.18% lower than native TSFlow and 3.56%–4.62% lower than TSFlow+Cal(0.5) across the four combinations.

Tables

Table C1. Context-only evaluation, observed-target Total (Total = MASE+ND+NRMSE; lower is better; three downstream initializations for a fixed generated training set per source).

Training sourceIII / GRUIII / TCNIV / GRUIV / TCN
Real2.1619±0.00102.1691±0.00252.1098±0.00172.1147±0.0014
Gaussian FM2.2153±0.00052.2188±0.00062.1586±0.00162.1643±0.0029
Gaussian + source calibration2.2136±0.00092.2199±0.00262.1673±0.00432.1711±0.0018
OU (α=0)2.2275±0.00452.2275±0.00092.1550±0.00362.1611±0.0026
OU-WSP (α=.03)2.1952±0.00202.2128±0.00262.1574±0.00332.1641±0.0045
Adapted TSFlow, native2.2697±0.00732.2895±0.00232.2505±0.01292.2585±0.0091
Adapted TSFlow + Cal(0.5)2.3016±0.01072.3065±0.00312.2370±0.00572.2587±0.0087

Table R1. Spectral comparison by dataset and generator seed (log-PSD discrepancy; lower is better).

DatasetSeedGaussianOU (α=0)OU-WSP
III20250.2415640.2251340.221712
III20260.2385380.2290030.203000
III20270.2539770.1576820.137773
IV20250.2050490.1610190.144422
IV20260.2372580.1625760.170334
IV20270.1932740.1493290.146831

Table R2. Matched-policy fidelity (256 contexts × 8 samples; lower is better).

DatasetSourceW1ACFLag1log-PSDBP viol.
IIIGaussian FM0.67080.01550.06650.24470.0897
IIIOU (α=0)0.64210.02840.06700.20390.0860
IIIOU-WSP (α=.03)0.71810.02920.06480.18750.0806
IIITSFlow native0.84130.03640.07370.40750.0926
IIITSFlow+Cal0.65870.03640.05980.24980.0771
IVGaussian FM0.64500.01760.05700.21190.1415
IVOU (α=0)0.63920.02800.05080.15760.1296
IVOU-WSP (α=.03)0.65530.02680.05040.15390.1305
IVTSFlow native0.99220.02920.07340.36760.1429
IVTSFlow+Cal0.82230.02920.05720.21510.1280
AuthorsOpenReview ↗

Thank you for the careful review and constructive guidance

Dear Reviewer XnMA,

Thank you for your careful reading and for the constructive feedback. We are grateful that you recognized the structured source design, the motivation for injecting temporal information from the observed context, and our explicit distinction between raw generation and deterministic BP repair.

Your two main concerns — whether the headline gains survive a strictly context-only pipeline, and how Table 2 should be interpreted under non-uniform post-processing — were central to our revision. They directly led us to run the full context-only GRU/TCN evaluation and the matched-policy fidelity analysis, and to state the evaluation boundary more clearly than in the original submission. Even though we did not have the chance to continue the discussion, your review shaped how we present the evidence and where we draw the claims.

Thank you again for your time and for helping us make the paper more rigorous. We will carry these points into the revision.

Sincerely, The authors

Reviewer_S9C6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose OU-WSP Flow Matching, a context-adaptive flow-matching framework for generating ICU time series. The method combines an Ornstein–Uhlenbeck (OU)-inspired autoregressive component with a Woodbury-spectral (WSP) component. Given an observed patient context, the model generates synthetic future time series, with additional post-processing to improve physiological consistency. Experiments on MIMIC-III and MIMIC-IV evaluate the generated data in terms of downstream forecasting utility, statistical fidelity, and physiological plausibility.

Critical strengths
  1. The paper focuses on an important and interesting problem of generating ICU time-series data. The authors demonstrate results on two real-world ICU datasets. The improvements demonstrate the potential utility for clinical EHR time series generation.

  2. The evaluation of the method is relatively comprehensive and transparent. The evaluation not only evaluates the downstream forecasting but also on the fidelity and physiological plausibility. This further confirms that the proposed method has potentials on real-world ICU. The additional ablation studies also provide useful insights into the behavior of different model components. Additionally, although there are some questions regarding the evaluation protocol (listed in later sections), I appreciate the transparent reports.

Critical weaknesses
  1. While the paper is positioned as an ICU-specific method, the challenges it identifies and addresses are not necessarily unique to ICU time-series data. For example, privacy and data accessibility are key motivations for the work; however, the proposed method still requires real patient contexts to generate synthetic futures. Moreover, the "physiology-aware" label in Figure 1 is not sufficiently justified. What specifically makes the proposed method physiology-aware? The constraint (SBP >= MBP >= DBP) is enforced through post-hoc processing, which could also be applied to other time-series generation methods. Overall, the connection between the ICU-specific challenges and the proposed method design remains underexplained. The authors could provide further justification.

  2. One of the core contributions is the WSP component introduced in Eq. (9). However, the experimental results show that its benefits are relatively marginal compared with the OU-only baseline, particularly in downstream forecasting utility. In the main forecasting experiments on MIMIC-III, the configuration disables the WSP residual entirely (\alpha = 0), while on MIMIC-IV, only small weights are used, yielding marginal improvements over the OU-only baseline. Although the additional experiments demonstrate some improvements in fidelity, these benefits are not consistent across datasets and evaluation metrics.

Questions for authors
  1. While the paper evaluates statistical fidelity and downstream forecasting utility, the diversity of generated samples remains insufficiently investigated. Could the authors provide a quantitative or qualitative evaluation of sample diversity across multiple generations and different patient contexts? Although Figure 2 shows variability across repeated generations for a single context, it does not sufficiently demonstrate the diversity of generated samples across different patient contexts.

  2. I am concerned about potential data leakage introduced by the preprocessing procedure. As mentioned in Appendix B, the original imputation procedure operates on the entire 48-hour window, potentially allowing future observations to influence the imputation of missing values in earlier context time steps. This is particularly problematic for time-series generation, where future values should be generated solely from past observations. Although Section 4.6 introduces a strictly causal preprocessing protocol, the main downstream forecasting results are not reevaluated under this setting. Could the authors clarify whether, and to what extent, this potential leakage affects the reported downstream forecasting performance?

3 comments
AuthorsOpenReview ↗

Response to Reviewer S9C6: Real-Context Generality, WSP Contribution, and Leakage Audit

We thank the reviewer for the thoughtful questions and clarify three points: real context/ICU specificity (Q1), WSP contribution (Q2), and diversity/leakage (W2).

Q1. Real patient context and ICU specificity

The model is conditional: it takes a real patient context and generates the future. Only one context window is needed per trajectory; the generated future can be fed back as the next context, so one window can seed continued generation without a large real cohort.

The generator is domain-agnostic: the same conditional flow-matching construction can be trained on other time-series domains. Its real use and adaptation depend on the training data, and the generation constraints (e.g., BP ordering) are domain-specific. ICU is our application setting, not a claim of a uniquely physiological mechanism; we do not claim unconditional synthesis or unrestricted release. We will validate cross-domain generalization in future work.

Q2. Contribution of the spectral residual and fidelity consistency

The residual is a context-selected Fourier correction to the OU anchor: frequencies come from the anchor residual, coefficients are fit on the context, and the correction is added to the endpoint. Its benefit is therefore concentrated in spectral and temporal statistics.

R1 shows lower log-PSD than Gaussian in all six seed combinations (23.4% on MIMIC-III, 27.4% on MIMIC-IV); against OU, the mean reductions are 8.1% and 2.4% (MIMIC-III 3/3 seeds, MIMIC-IV 2/3). Matched-policy fidelity shows the expected metric dependence: OU-WSP has the lowest log-PSD on both datasets (0.1875/0.1539) and the lowest Lag1/BP violation on MIMIC-III, while OU has the lowest W1, Gaussian the lowest ACF, and TSFlow+Cal the lowest BP violation on MIMIC-IV. This is a spectral-versus-marginal trade-off, not a failure; WSP is not a universal metric winner.

W1. Diversity across patients

Diversity. R4 reports 256 held-out contexts with eight samples each. OU-WSP has higher between-context spread and spread ratio than Gaussian on both datasets (0.689 vs 0.681 and 0.810 vs 0.800; 0.787 vs 0.722 and 0.915 vs 0.840), but lower within-context spread (0.576 vs 0.601; 0.629 vs 0.667). These descriptors support variability, not coverage superiority or complete mode coverage. TSFlow diversity is omitted because its sampling RNG seed was unrecorded.

W2.Leakage impact

Leakage. All generator fits and synthetic training contexts use training-split windows only; validation and test contexts are never used to fit generators or construct synthetic training data. The audit identified 226/26/57 affected train/validation/test windows in MIMIC-III and 519/88/157 in MIMIC-IV; we rebuilt both datasets with context-only imputation and retrained the generators. C1 is the corrected evaluation. A leakage-only delta is not identifiable because the original pipeline was not rerun under the same protocol: the 22.0%–27.2% comparison and C1 differ in preprocessing and training, so we report corrected numbers and window counts, not a spurious effect.

Tables

Table C1. Context-only evaluation, observed-target Total (Total = MASE+ND+NRMSE; lower is better; three downstream initializations for a fixed generated training set per source).

Training sourceIII / GRUIII / TCNIV / GRUIV / TCN
Real2.1619±0.00102.1691±0.00252.1098±0.00172.1147±0.0014
Gaussian FM2.2153±0.00052.2188±0.00062.1586±0.00162.1643±0.0029
Gaussian + source calibration2.2136±0.00092.2199±0.00262.1673±0.00432.1711±0.0018
OU (α=0)2.2275±0.00452.2275±0.00092.1550±0.00362.1611±0.0026
OU-WSP (α=.03)2.1952±0.00202.2128±0.00262.1574±0.00332.1641±0.0045
Adapted TSFlow, native2.2697±0.00732.2895±0.00232.2505±0.01292.2585±0.0091
Adapted TSFlow + Cal(0.5)2.3016±0.01072.3065±0.00312.2370±0.00572.2587±0.0087

Table R1. Spectral comparison by dataset and generator seed (log-PSD discrepancy; lower is better).

DatasetSeedGaussianOU (α=0)OU-WSP
III20250.2415640.2251340.221712
III20260.2385380.2290030.203000
III20270.2539770.1576820.137773
IV20250.2050490.1610190.144422
IV20260.2372580.1625760.170334
IV20270.1932740.1493290.146831

Table R4. Diversity descriptors (256 contexts × 8 samples; three-generator-seed means; normalizer units).

DatasetSourcewithinbetweenspread ratio
IIIGaussian FM0.601±0.0110.681±0.0130.800±0.015
IIIOU-WSP0.576±0.0280.689±0.0050.810±0.006
IVGaussian FM0.667±0.0010.722±0.0090.840±0.010
IVOU-WSP0.629±0.0060.787±0.0180.915±0.020

Thank you for the authors' response. But I still think some my main concerns are not resolved. Under the leakage-free evaluation, OU-WSP’s advantage over the Gaussian flow-matching baseline seems to disappear. The new results also confirm that the WSP component contributes little. In addition, the protocols behind the new results are not clearly stated, and the response effectively concedes that the method is neither ICU- nor physiology-specific. Addressing these issues would require large effort in revising the contributions and motivations, and re-running the main experiments under a consistent leakage-free protocol. Therefore, I decide to maintain my score.

AuthorsOpenReview ↗

Thank you for the careful review and guidance

Dear Reviewer S9C6,

Thank you for your careful reading and for the constructive feedback throughout the review process. We are grateful that you recognized the importance of ICU time-series generation, the value of evaluating on both MIMIC-III and MIMIC-IV, and the transparency of our evaluation across downstream utility, statistical fidelity, and physiological plausibility.

We also appreciate your candid concerns — the ICU/physiology-specific framing, the marginal benefit of the WSP residual, the diversity analysis, and the need for a consistent leakage-free protocol. Even though the rebuttal did not fully resolve them to your satisfaction, these points have made it clear where the paper's claims need to be tightened, and they will directly shape the revision: a clearer scope statement, a more careful characterization of the residual's spectral contribution, and a consistent context-only evaluation.

Thank you again for your time and for helping us see the paper more critically. We will make the revision more precise and more honest because of your review.

Sincerely, The authors

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

This paper proposes OU-WSP FM, a conditional flow-matching framework for ICU time-series generation. The method constructs a context-adaptive source distribution by combining an Ornstein–Uhlenbeck (OU) autoregressive prior with Fourier-based spectral residual extrapolation. Using a Transformer-based flow-matching model, the authors evaluate the approach on MIMIC-III and MIMIC-IV, reporting improvements in downstream forecasting utility and several statistical fidelity metrics over Gaussian FM.

Critical strengths
  1. The paper addresses ICU time-series generation, considering both downstream forecasting utility and statistical fidelity.

  2. The combination of OU dynamics and spectral residual modeling provides an intuitive way to incorporate temporal structure into the source distribution.

Critical weaknesses
  1. Limited methodological novelty: TSFlow (Kollovieh et al., 2025) has already explored replacing the standard Gaussian source distribution in flow matching with temporally structured priors, including OU-based priors. Fourier-based temporal modeling and extrapolation are also well established.

  2. Insufficient validation of the proposed source distribution. Table 4 shows that OU-WSP overall underperforms OU-only on W1 across both datasets and output calibration settings, while its log-PSD improvement is limited to MIMIC-III. Despite improvements in cross-channel correlation metrics, the results do not establish a consistent overall advantage. Moreover, the evaluation uses only a single random seed and does not include downstream forecasting. Experiments fail to demonstrate the effectiveness and generalizability of OU-WSP.

  3. Potential information leakage in the main experiments. Appendix B.1 acknowledges that the original preprocessing pipeline used complete 48-hour windows for imputation, potentially introducing future information into the observed contexts. Although Sec. 4.6 introduces a strictly causal protocol, the downstream forecasting experiments in Table 1 are not repeated. Therefore, the reported forecasting improvements remain insufficiently validated.

  4. Incomplete and potentially unfair baseline comparisons: The paper lacks a direct experimental comparison with TSFlow, the most closely related method. Moreover, Table 2 uses inconsistent or incompletely documented post-processing procedures across baselines, while some baseline implementations are adapted rather than official reproductions. These issues weaken the reliability of the reported comparisons.

  5. Limited modeling of cross-channel dependencies and physiological constraints. Although the model shares parameters across channels, each channel is generated separately, limiting explicit modeling of their joint conditional distribution. Furthermore, blood-pressure ordering is enforced through deterministic post-processing rather than learned by the generator. The claims regarding physiological consistency should therefore be more carefully qualified.

Questions for authors

Please refer to the Weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer YBax: Mathematical Source Construction, Context-Only Evaluation, and TSFlow Comparison

We thank the reviewer. We completed the context-only evaluation, a three-generator-seed spectral analysis, and a released-TSFlow comparison.

W1. Methodological contribution

Our contribution is a context-adaptive residual-spectral source: form rt=xt−x^tOUr_t=x_t-\hat x^{OU}_t; RFFT selects top-QQ bins FQF_Q. Fit β^=arg⁡min⁡β∑t∣rt−∑f∈FQβfei2πft/L∣2\hat\beta=\arg\min_\beta\sum_t|r_t-\sum_{f\in F_Q}\beta_f e^{i2\pi ft/L}|^2; extrapolate r^L+h=∑f∈FQβ^fei2πf(L+h)/L\hat r_{L+h}=\sum_{f\in F_Q}\hat\beta_f e^{i2\pi f(L+h)/L}; set x^L+h=x^L+hOU+αr^L+h\hat x_{L+h}=\hat x^{OU}_{L+h}+\alpha\hat r_{L+h}. For fixed FQF_Q, this predictor is the diffuse-prior limit of a finite-feature Gaussian posterior mean. The context-selected residual basis distinguishes it from TSFlow; no dense inversion.

W2. Validation beyond the single-seed diagnostic

  • C1 (context-only downstream). Each source has 6,000 pairs; real-side scoring uses rebuilt data and validation-selected checkpoints. OU-WSP achieves the lowest observed Total among synthetic sources on MIMIC-III (2.1952 GRU, 2.2128 TCN); it is 0.01%–0.91% below Gaussian and 1.54%–2.34% above Real.
  • R1 (spectral, three seeds). OU-WSP has lower log-PSD than Gaussian in all six combinations (mean reductions 23.4%/27.4%); against OU, 8.1%/2.4% (MIMIC-III 3/3, MIMIC-IV 2/3).
  • R2 (matched fidelity). With no BP repair/filtering, OU-WSP has the lowest log-PSD on both datasets (0.1875 vs 0.2447; 0.1539 vs 0.2119). R2 is metric-specific: OU has the lowest W1, Gaussian the lowest ACF, and OU-WSP the lowest log-PSD and BP violation.

W3. Information boundary

All generator fits and synthetic training contexts use training-split windows only; validation and test contexts are never used to fit generators or construct synthetic training data. The audit identified 226/26/57 affected train/validation/test windows in MIMIC-III and 519/88/157 in MIMIC-IV; these were rebuilt with context-only imputation and the generators retrained, and C1 is the resulting evaluation. The 22.0%–27.2% comparison used a different protocol; retained as history.

W4. Baselines and comparison fairness

TSFlow. We adapted released TSFlow to 24-hour ICU contexts (context_freqs: 14 → 1, use_lags: False), keeping its S4 backbone, OU kernel, Euler steps, EMA, batch 64. At 51,200 updates, OU-WSP is 3.28%–4.18% lower than native TSFlow and 3.56%–4.62% lower than TSFlow+Cal; whole-model comparison (187k vs 21M parameters), one TSFlow seed.

Fidelity policies. R2 evaluates all methods without BP repair/filtering and states calibration per row; the submitted Gaussian row is daggered as BP-filtered.

Other implementations. TimeGAN, RGAN, TimeVAE, COT-GAN and Diffusion-TS are adaptations; released TSFlow is the closest structured-prior baseline.

W5. Cross-channel relationships and clinical framing

Theoretical lens. Ren and Sutherland's decomposition gives Δlog⁡πt(y∣xo)=−ηAtKtGt+O(η2)\Delta\log\pi^t(y|x_o)=-\eta A^t K^t G^t+O(\eta^2) for one step on (xu,yu)(x_u,y_u), with A=I−1π⊤A=I-\mathbf{1}\pi^\top the softmax Jacobian, KK the eNTK, and GG the logit gradient. Under their stable-eNTK assumption, changes accumulate as kernel-weighted sums. In our shared backbone KK couples channels; the source carries part of the target's spectral energy, reducing GG's projection onto high-frequency, small-eigenvalue modes. A linearized mode-kk error then decays as (1−ηλk)t(1-\eta\lambda_k)^t, the learning-dynamics reading of the log-PSD reduction. This is a heuristic, not a full 20M-SiT theorem; BP ordering is post-hoc.

Tables

Table C1. Context-only evaluation, observed-target Total (Total = MASE+ND+NRMSE; lower is better; three downstream initializations for a fixed generated training set per source).

Training sourceIII / GRUIII / TCNIV / GRUIV / TCN
Real2.1619±0.00102.1691±0.00252.1098±0.00172.1147±0.0014
Gaussian FM2.2153±0.00052.2188±0.00062.1586±0.00162.1643±0.0029
Gaussian + source calibration2.2136±0.00092.2199±0.00262.1673±0.00432.1711±0.0018
OU (α=0)2.2275±0.00452.2275±0.00092.1550±0.00362.1611±0.0026
OU-WSP (α=.03)2.1952±0.00202.2128±0.00262.1574±0.00332.1641±0.0045
Adapted TSFlow, native2.2697±0.00732.2895±0.00232.2505±0.01292.2585±0.0091
Adapted TSFlow + Cal(0.5)2.3016±0.01072.3065±0.00312.2370±0.00572.2587±0.0087

Table R1. Spectral comparison by dataset and generator seed (log-PSD discrepancy; lower is better).

DatasetSeedGaussianOU (α=0)OU-WSP
III20250.2415640.2251340.221712
III20260.2385380.2290030.203000
III20270.2539770.1576820.137773
IV20250.2050490.1610190.144422
IV20260.2372580.1625760.170334
IV20270.1932740.1493290.146831
AuthorsOpenReview ↗

Thank you for the rigorous review and guidance

Dear Reviewer YBax,

Thank you for your careful and demanding review. We are grateful that you recognized the importance of ICU time-series generation and the intuitive value of combining OU dynamics with spectral residual modeling.

Your criticisms were especially helpful in sharpening the paper: the need to position the contribution relative to TSFlow in precise mathematical terms, to distinguish source construction from baseline adaptation, to re-examine the preprocessing boundary and retrain under a context-only protocol, and to qualify claims about cross-channel dependencies and BP post-processing. These points directly shaped our rebuttal and will shape the revision — a clearer mathematical framing of the residual construction, a leakage-aware evaluation, a direct TSFlow comparison, and more careful language around physiological consistency.

Thank you for the time and rigor you invested. Your review made the paper's claims more precise and its limitations more honest, and we will carry that into the revision.

Sincerely, The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.