acceptodds
Under review as a conference paper at ICLR 2027

Distribution Recovery with Domain Constraints: Applications to Medical Time Series

Abstract

Recovering the occurrence and severity of physiological abnormalities from intermittent clinical measurements requires a distribution over possible trajectories, yet sparse observations may not uniquely determine this distribution. SFBD-OMNI guides recovery with clean reference sample anchors, which are generally unavailable in retrospective medical records. In this work, we introduce Projected SFBD, which directly incorporates domain knowledge as constraints on recovered trajectories through a projection-based update. We provide theoretical convergence guarantees and show how physiological priors narrow the space of admissible distributions and can improve identifiability without clean reference samples. On irregular MIMIC-III charts, evaluation against bedside-monitor references shows better trajectory-distribution scores and more accurate recovery of hypotension occurrence, severity, recurrence, and within-stay variation in vital signs than strong baselines. Ablations on MIMIC-III and the Pulse simulator show that projection better preserves clinically relevant variation and yields less biased clinical summaries than constructed clean anchors.

Primary area:
learning on time series and dynamical systems
Keywords:
Medical time series, Distribution recovery, Diffusion models

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

This paper proposes Projected SFBD, which incorporates physiological domain constraints into the projection updates of distribution recovery, enabling recovery of the latent trajectory distribution of sparse, irregular medical time series without complete reference samples. The authors provide a conditional convergence guarantee and, through experiments on MIMIC-III and the Pulse simulator, demonstrate the method’s advantages in distribution recovery as well as in hypotension event detection, severity estimation, and recurrence count estimation.

Critical strengths
  1. The paper addresses distribution recovery under sparse medical observations and leverages domain knowledge to reduce reliance on complete reference samples, which offers clear practical value.

  2. Beyond distributional metrics such as Energy Score and Variogram Score, the paper evaluates clinically relevant tasks, including hypotension detection, severity estimation, and event recurrence counts, and validates the method’s effectiveness through ablation studies.

Critical weaknesses
  1. Please further show that the practical recording-condition projection satisfies Assumption 2, and clarify to what extent the theoretical convergence guarantee applies to the implemented algorithm.

  2. I suggest comparing performance with and without the projection while keeping prediction positions, architecture, training budget, and noise settings identical, to establish whether the constraint projection is the main source of improvement.

  3. It would be helpful to demonstrate, through theoretical analysis or controlled experiments, that the proposed physiological constraints genuinely improve distributional identifiability, rather than merely shrinking the feasible trajectory space.

I would consider raising my score if these concerns are adequately addressed.

Questions for authors

See Weaknesses.

2 comments
AuthorsOpenReview ↗

Thank you for the comments

Thank you for your constructive comments. We address the points raised below and provide the additional results requested.

Q1. Please further show that the practical recording-condition projection satisfies Assumption 2.

A: Assumption 2 requires that projection leave the true distribution unchanged. Our implementation satisfies this whenever true trajectories obey the constraints. Notably, these constraints are not modelling choices that may or may not hold; they encode established physiology. The ordering DBP ≤ MAP ≤ SBP follows from the definition of mean arterial pressure as the time average of arterial pressure over the cardiac cycle, which necessarily lies between its minimum and maximum, and channel bounds such as SpO$_2 \le 100%$ are physical limits. True trajectories therefore satisfy them, and Assumption 2 holds exactly for these constraints. Consistent with this, on the Pulse simulator the true trajectories never violate the BP ordering (0.000% of cells). In an audit of MIMIC-III charts, only 0.29% of co-observed pairs violate it, mostly due to timing and aggregation artifacts; these recorded values remain unchanged. The MAP band and the minimum pulse pressure encode the standard clinical approximation MAP ≈ DBP + (SBP − DBP)/3 and a physiological lower limit on pulse pressure. We impose them with tolerances calibrated on reference data, which absorb measurement noise, so Assumption 2 holds approximately for these two, with violation rates of at most about 1% on arterial measurements: 0.68% for the pulse-pressure floor, and the MAP tolerance is set at the 99th percentile of the arterial residual (Appendix C.2).

Q2. To what extent the theoretical convergence guarantee applies to the implemented algorithm.

A: Regarding the scope of the guarantee, Theorem 1 applies to the implemented update rule, including the adaptive replacement fraction, skip rule, and projection. The only departure from the theory is the error signal: Eq. 7 uses a difference of KL divergences, which requires densities and cannot be estimated directly from samples, whereas the implementation uses the corresponding difference of MMD$^2$ values, which admits an unbiased sample estimator. This substitution preserves the essential convergence structure: with a characteristic kernel, MMD vanishes if and only if the compared distributions coincide, exactly when the KL criterion vanishes, so the two criteria share the same optimum. (Also see the newly added discussion at the end of sec 4.2.) The update rule also keeps the accepted MMD errors summable, mirroring the condition in Theorem 1. To verify convergence empirically, we added Appendix D.6. Theorem 1 guarantees that trajectories drawn from the learned distribution, when recorded in the same way as real charts, reproduce the distribution of the observation records. We test this on records the model never sees: about 15% of each stay’s chart values are held out from training, and we measure how well the model predicts them using CRPS. As shown in Appendix D.6, held-out CRPS decreases steadily, from 0.504 to 0.206 for vital signs and from 0.530 to 0.268 for laboratory values, showing that the fit to the observation records keeps improving, as the theorem predicts. CRPS does not reach zero because the held-out values are not fully determined by the remaining observations.

Q3. I suggest comparing performance with and without the projection while keeping prediction positions, architecture, training budget, and noise settings identical, to establish whether the constraint projection is the main source of improvement.

A: We have added SFBD without projection to Tables 1 and 2, keeping the prediction locations, architecture, training budget, and noise settings identical to isolate the effect of projection. Projection improves every score: ES and VS decrease across all window lengths (e.g., 1 h ES: 0.920 → 0.903; 6 h ES: 2.470 → 2.430), while all downstream metrics improve, including AUROC (0.926 → 0.935), sensitivity at 90% specificity (78.6% → 81.9%), severity bias (−3.17 → −0.61 mmHg), severity MAE (7.07 → 5.63), and episode-count MAE (1.64 → 1.58). The matched ablation in Table 7 further confirms these gains, with CRPS decreasing from 0.227 to 0.216 and coverage increasing from 0.772 to 0.832. While the SFBD framework accounts for much of the improvement over external baselines, projection provides consistent additional gains, particularly in estimating the severity of hypotensive events, where blood-pressure constraints are most relevant.

Q4. It would be helpful to demonstrate, through theoretical analysis or controlled experiments, that the proposed physiological constraints genuinely improve distributional identifiability, rather than merely shrinking the feasible trajectory space.

A: We have added a toy example immediately after Definition 1. It shows two distributions that produce identical records and are therefore indistinguishable from the data, although every individual value in both is physiologically plausible. A physiological constraint rules out the distribution that pairs values in an impossible way (MAP above SBP) and keeps the true one. Identifiability is thus restored not by narrowing the range of individual values, which are all plausible, but by excluding combinations of values that cannot occur together.

Empirically, the with/without-projection comparison in Table 1 (see our response to Q3) supports the same mechanism. Both variants are trained on the same charts, and their outputs are compared directly against bedside-monitor references, which serve as ground truth and are never used in training. Projected SFBD achieves better scores at every window length, indicating that projection learns a distribution closer to the truth and providing indirect evidence that the constraints improve identifiability in practice. The Pulse results against known truth (Table 4) point in the same direction.

Reviewer_boU1OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper extends SFBD-OMNI to recover the distribution of complete ICU trajectories from sparse, irregular chart data. Instead of using clean anchor samples, each sampled completion is projected onto a set of physiological constraints (e.g. range bounds, DBP ≤ MAP ≤ SBP). The authors provide a KL convergence result, and evaluate on MIMIC-III against bedside-monitor data, reporting better scoring rules and hypotension detection than CSDI, SAITS, linear interpolation and forward fill.

Using monitor data as a reference inside chart gaps is a nice evaluation design. However, I am afraid the gains cannot really be attributed to the proposed projection, and the comparison with baselines is not that fair.

Critical strengths
  • The evaluation against bedside-monitor data within chart gaps is clinically meaningful, and the event-level metrics (e.g. hypotension detection) are relevant.
  • The constraint projection is simple, and the blood-pressure constraints make sense clinically (appreciate!).
Critical weaknesses

I am mainly concerned that the gains do not come from the projection, which is the main contribution, and that the baselines are not compared fairly.

Clarity and Presentation

  • The method appears under three different names ("Projected SFBD", "Prior-Constrained SFBD-OMNI" and "Constrained SFBD"), which is a bit confusing.
  • The Pulse simulator experiment is not described (cohort, sparsification, seeds), and there is no limitations section.

Methodology

  • The theory is a fairly incremental extension of SFBD-OMNI and SFBD Flow. As far as I can see, Theorem 1 also holds when the projection is the identity, so it does not really show that the projection helps.
  • The paper claims "weak assumptions" in L52, but the assumptions look rather strong to me, e.g. summable posterior errors against the unknown true posterior, and a target-preserving projection. The authors also acknowledge these are not verified empirically.
  • Identifiability is the main motivation for replacing anchors with constraints, but I failed to see any example (toy or real) where the constraints actually make the problem identifiable.

Experiment and Results

  • In Table 7, projection only changes CRPS from 0.229 to 0.227. In Table 3, SFBD with constructed anchors and no projection already reaches AUROC 0.932, against 0.935 with projection and 0.847 for CSDI. So the gains seem to come mostly from the SFBD pipeline rather than the constraints? An "SFBD without projection" row in Tables 1 and 2 would be expected.
  • CSDI and SAITS are given hourly-binned inputs, while SFBD keeps continuous timestamps. This is not a fair comparison, I am afraid. Linear interpolation even beats both in hypotension AUROC, which suggests they may not be well tuned.

Minor Issues and Suggestions

  • The cohort size is reported as both 5,686 and 5,486 stays (L899, L1196).
Questions for authors

I would expect some clarification on:

  1. What are the Table 1 and 2 results for SFBD without projection, under the same training schedule and query points?
  2. How do CSDI and SAITS perform at native or 5-minute resolution?
  3. Is there a concrete case where the constraints restore identifiability?
2 comments
AuthorsOpenReview ↗

Thank you for the comments

Thank you for your comments. Here are our responses.

Q1. Regarding the effectiveness of the physiological constraints and the comparison to the version without projection.

A: We agree that a substantial part of the gain over external baselines comes from the SFBD framework itself, and we have added SFBD without projection to Tables 1 and 2 to separate the two. Projection improves on it consistently: ES and VS decrease at every window length (e.g., 1 h ES 0.920 → 0.903), and every downstream metric improves, including AUROC (0.926 → 0.935), sensitivity at 90% specificity (78.6% → 81.9%), severity bias (−3.17 → −0.61 mmHg) and severity MAE (7.07 → 5.63). The gains are largest on the severity of hypotensive events, where the blood-pressure constraints are active.

Regarding Table 3, AUROC alone does not capture the effect of constructed anchors. While their AUROC is close to that of Projected SFBD (0.932 vs 0.935), they reduce coverage from 0.832 to 0.608 and underestimate the episode count by 40%, against −4% with projection. Constructed anchors therefore preserve detection while distorting the trajectory distribution, which projection avoids.

Q2. Regarding a better comparison to the baseline and a better implementation of CSDI and SAITS.

A: We thank the reviewer for raising this point. In the original submission, CSDI and SAITS were trained and evaluated on hourly-binned inputs, while the evaluation itself is performed at 5-minute resolution against bedside-monitor references, which disadvantaged both methods. We have retrained them on a native 5-minute grid and updated Tables 1 and 2 accordingly. Even against these stronger baselines, Projected SFBD remains the best on every score and by a clear margin on distributional and episode-level metrics: relative to the strongest baseline in each row, it lowers ES by 11–13% and VS by 14–23% across window lengths, reduces the episode-count bias from −49.2% (SAITS) to −4.5%, and lowers the per-stay count MAE from 2.20 to 1.58. SAITS still has a low AUROC (0.696) for a structural reason rather than tuning: it returns a single deterministic completion, so the estimated probability of a hypotensive episode is either 0 or 1, which limits its ability to rank stays.

Q3. Regarding the statement presented in Thm 1 and its novelty.

A: We acknowledge that the convergence result in Theorem 1 also holds without projection. This is intentional: the theorem establishes that projection preserves the convergence guarantees of SFBD-OMNI and SFBD Flow, while its benefit lies in identifiability. By restricting every iterate to the physiologically admissible family, projection narrows the set of possible limiting distributions, potentially enabling recovery of the true distribution without clean anchors (Definition 1 and the subsequent toy example). Empirically, the added comparisons with SFBD without projection in Tables 1 and 2 demonstrate improvements in distribution-level scores.

Q4: Regarding the paper's claim of "weak assumptions" in L52 like summable posterior errors against the unknown true posterior, and a target-preserving projection.

A: We respectfully believe that these assumptions are weak. Summable posterior errors is essentially a property we construct rather than assume: the skip rule accepts an update only when its discrepancy falls below the recent average, which is designed to make the accepted errors decay and keep their sum finite (Section C.2), following the same mechanism as SFBD Flow. It does not require the learned posterior to be exact, only that updates which do not improve it are rejected. The only change from the theory is the error measure: the KL difference in Eq. 7 cannot be computed from samples, so we replace it with the corresponding MMD difference, and the two criteria share the same optimum, since the MMD vanishes if and only if the KL divergence does. Target preservation is likewise not an additional technical requirement but follows directly from the premise of the method: if the physiological constraints hold for real patients, true trajectories are feasible and the projection leaves them unchanged (Assumption 2). Many such constraints are available in medicine, from physical bounds and the ordering of blood pressures to acid–base and hemoglobin–hematocrit relations, which makes the approach broadly applicable. We have made corresponding changes in the revised version to clarify our points.

Q5: Regarding the examples for solving identifiability problems using projection

A: We have added an example immediately after Definition 1 showing how a physiological constraint makes the problem identifiable: two trajectory distributions that are indistinguishable from the records are separated once the infeasible one is excluded. We have also added results for SFBD without projection to Tables 1 and 2. Projection consistently improves the distribution-level scores, which is consistent with the constraints narrowing the set of admissible distributions on real data.

Q6: Regarding the Pulse simulator experiment spec.

A: We have added the Pulse Simulator configuration in Appendix C.5.

Q7: Regarding other issues and typos

A: We have fixed all the editorial issues per your suggestion.

Reviewer_7odhOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Projected SFBD to extend SFBD-OMNI to recover time-series distributions from incomplete observations. The method incorporates physiological bounds and cross-channel constraints through projection and provides theoretical analysis of convergence and identifiability under additional assumptions. Experiments on MIMIC-III and simulated data compare the method with existing baselines on distributional recovery and downstream clinical measures.

Critical strengths

The paper incorporates physiological constraints into SFBD-OMNI to recover clinical time series from incomplete observations. It also provides theoretical analysis and evaluates the method on real ICU data using bedside-monitor measurements as references.

Critical weaknesses

The key concerns are the effectiveness of the physiological constraints, the gap between the theoretical convergence guarantee and the practical implementation, and the lack of an independent test cohort. Please see the questions to the authors for more details.

Questions for authors
  1. Identifiability concern. Is it possible that many different trajectories satisfy the physiological constraints while matching the same observations? Also, how does the method handle noisy or inconsistent measurements that violate these constraints?
  2. In Eq.13-15, projection may also change their distribution. How does the method ensure that projection does not introduce additional bias?
  3. Convergence concern. How is the MMD-based criterion in Eq.17 related to the theorem 1?
  4. Missingness. Section 4.1 says observation times can depend on patient trajectories, but Section 5 uses independent re-corruption with channel-specific rates. Does this capture informative missingness in ICU data?
  5. Evaluation. Section C.1 states that the 669 waveform-matched ICU stays were also used for training, and the same monitor references were used for configuration selection. How can this affect the reported results? Could the authors evaluate the method on an independent test cohort?
2 comments
AuthorsOpenReview ↗

Thank you for the comments

Thank you very much for your positive and constructive comments. Here are our responses.

Q1. Regarding the Identifiability: Is it possible that many different trajectories ...

A: This is possible in general, as physiological constraints do not guarantee unique recovery. Instead, they narrow the set of candidate distributions. Without constraints, any distribution reproducing the observed records is equally valid. With constraints, distributions assigning mass to physiologically impossible trajectories are excluded, while the true distribution is retained. Identifiability holds when only the true distribution remains (Definition 1).

For noisy measurements, approximate physiological relations are imposed as tolerance bands rather than exact equalities, e.g., MAP is only required to lie within a calibrated band around (SBP + 2 DBP)/3 (Appendix C.2), which absorbs measurement noise. For inconsistent measurements, projection never modifies recorded values; constraints apply only to generated values.

Q2. Projection bias related to Eqs. 13–15.

A: The projection does not introduce bias because the constraints are established physiological relations, and true patient trajectories already satisfy them. The projection therefore leaves the true distribution unchanged. It only corrects generated trajectories that are physiologically impossible and so cannot be the truth. Formally, the projection can only move the estimate closer to the true distribution, never farther from it.

The purpose of the projection is not to reshape the distribution. Without clean reference samples, many trajectory distributions fit the incomplete records equally well. Applying the projection at every iteration removes the physiologically impossible candidates while keeping the true distribution. This narrows the hypothesis space and improves identifiability (Definition 1).

Q3. Convergence: How is the MMD-based criterion in Eq. 17 related to Theorem 1?

A: Eq. 17 is the practical implementation of the adaptive rule in Eq. 6 that Theorem 1 analyzes. The only change is the error signal. Theorem 1 uses the posterior-error bound δk\delta_k of Eq. 7, a difference of KL divergences, which cannot be computed from samples. The implementation replaces it with the corresponding difference of MMD$^2$ values: the discrepancy between proposed and stored completions minus that between two stored subsets, which removes sampling noise. The two criteria share the same optimum: both vanish exactly when the compared distributions coincide. Although the proxy does not certify Eq. 7, the update rule is designed to keep the sum of the accepted proxy errors finite, mirroring the summability condition in Theorem 1. Thus, Eq. 17 provides a practical convergence criterion aligned with the recovery objective rather than merely a heuristic, and Appendix D.6 confirms this empirically by showing that held-out accuracy improves steadily over training. We have improved the writing in the corresponding part at the end of Sec. 4.2 to avoid any confusion.

Q4. Missingness: Section 4.1 states that observation times can depend on patient trajectories, but Section 5 uses independent re-corruption with channel-specific rates. Does this capture informative missingness in ICU data?

A: While our theory covers trajectory-dependent observation times (Section 4.1, Theorem 1), our empirical study showed that modelling this dependence in re-corruption does not meaningfully improve results. Replacing rate-matched masking with a learned value-dependent kernel changes ES and VS by only 0.2–3%, comparable to cross-run variation, and performs worse on all downstream metrics except range error, even though the kernel captures informative charting, notably for lactate. We therefore adopted the simpler rate-matched operator in the experiments. We have now added this ablation as Appendix D.4.

Q5. The effectiveness of the physiological constraints

A: We have added the numbers for the variant without projection in Tables 1 and 2. The results show that the projection is most useful where sparse charts leave clinically important quantities undetermined and physiological constraints are informative about them, most notably the severity and occurrence of hypotensive events. This matters clinically because treatment decisions depend on abnormal events, which often occur between chart measurements.

Q6. For the lack of an independent test cohort.

A: We have added a new appendix D3 for the results on an independent test cohort whose references were never used during development. The conclusions there match those in the main text, showing that they do not depend on reusing these references for configuration selection.

Reviewer_7odhOpenReview ↗

Maintain original rating

Thank you for the clarifications.I appreciate the authors' efforts to address my concerns. I believe some theoretical concerns regarding projection bias and the connection between the practical convergence criterion and the theoretical guarantees remain insufficiently resolved. Therefore, I will maintain my original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.