acceptodds
Under review as a conference paper at ICLR 2027

Neural Belief-State Filtering with Cross-Variable Coupling for Irregular Time Series

Abstract

While many sequence models assume evenly spaced observations, in domains such as intensive care different measurements and interventions occur at irregular, asynchronous times, and the underlying state is only partially observed. We introduce Neural Belief-State Filtering (NBSF), a continuous-discrete filter that maintains a probabilistic belief over observed and latent variables, propagates it across irregular gaps with a structured exponential-decay backbone and a gap-dependent nonlinear residual, and assimilates each observation through a learned, innovation-driven update. The distinguishing feature of NBSF is that an observation of one variable can directly revise the belief about other, currently unobserved variables at event time rather than only influencing them later through temporal propagation and it does so without maintaining a dense filtering covariance. We study three assimilation variants: a local-coordinate control that updates only the observed variable, a learned dense cross-variable gain, and a nonlinear cross-variable update. Across four clinical benchmarks and a controlled sensing task, event-time cross-variable assimilation is most useful when direct target observations are scarce, and its advantage grows under target suppression; the dense-gain variant provides the most consistent benefit, while the nonlinear update shows dataset-dependent gains. Parameter-matched ablations show these gains are not explained by model size. Our central finding is that event-time cross-variable information flow is the more consistent mechanism under target scarcity, whereas additional nonlinear update complexity is not consistently beneficial. NBSF uses only 140K–170K parameters, provides an explicit probabilistic predictive belief whose calibration is evaluated in-distribution, and achieves competitive performance across established clinical irregular-time-series benchmarks.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Neural Belief-State Filtering, irregular time series, multivariate time series, cross-variable assimilation, probabilistic filtering, continuous-discrete state-space models, missing data, clinical time series, uncertainty quantification, belief-state estimation, target dropout, multivariate forecasting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_5sejOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Neural Belief-State Filtering (NBSF), a continuous-discrete filtering framework for irregular multivariate time series. Instead of maintaining only a deterministic hidden state, NBSF represents the evolving state as a probabilistic belief and separates temporal propagation from observation-time assimilation. Its main contribution is an event-time cross-variable update, allowing an observation of one variable to immediately revise beliefs about other unobserved variables without maintaining a dense covariance matrix. The paper evaluates local-coordinate, dense-gain, and nonlinear cross-variable updates across several clinical datasets and a controlled sensing dataset, with the main finding that cross-variable assimilation is particularly useful when direct observations of the target variable are scarce.

Critical strengths

The strongest part of the paper is the mechanism-level evaluation. The authors do not rely only on overall classification performance; they test belief quality before direct observation, progressively remove target observations, compare against a parameter-matched wide-local model, and also destroy cross-variable structure by permuting non-target variables across patients. The latter is especially convincing because the advantage of the cross-variable models disappears and eventually reverses as cross-variable information is destroyed, which directly supports the claimed mechanism rather than merely showing a larger model performs better.

Critical weaknesses

My main reservation is that the strongest evidence supports event-time cross-variable information transfer, but not necessarily the full nonlinear NBSF update proposed as the main model. In several experiments, the much simpler dense-gain variant is as good as or better than the nonlinear cross-variable update. Dense-gain gives the lowest CRPS in all four core belief-quality comparisons, performs more consistently under target suppression, and is best in most subsystem-holdout settings. In addition, the propagation ablation shows that the hybrid propagation mechanism is not consistently better than the simpler SSM-only or MLP alternatives, and removing the latent block actually slightly improves P19 classification. These results do not weaken the central cross-variable-assimilation idea, but they do suggest that some architectural components of the full NBSF formulation are less strongly justified than the paper’s broader method presentation may imply.

Questions for authors

The main question I would like the authors to clarify is what they consider the essential contribution of NBSF after the ablation results. The experiments seem to provide strong evidence for event-time cross-variable assimilation, but considerably weaker evidence that the nonlinear update, latent state, or hybrid propagation are necessary. Since dense-gain is frequently the strongest variant, would the authors view the simpler dense-gain filter as the preferred realization of the proposed principle? I think making this distinction clearer would strengthen the paper, because the empirical contribution appears to be more about where cross-variable information enters the filtering process than about the complexity of the neural update itself.

4 comments

We thank the reviewer for recognizing the mechanism-level experiments. We agree that NBSF’s essential filtering principle should be distinguished from optional architectural complexity.

1. What is the essential contribution?

NBSF propagates a factorized Gaussian belief across irregular gaps and assimilates measurements through learned, innovation-driven corrections, with O(S)O(S) filtering-state storage. The main question is when an observation should immediately revise beliefs about unobserved variables.

Cross-variable updates already exist in Kalman filtering; we do not claim otherwise. Our contribution is a learned continuous-discrete realization and controlled tests of the jointly learned observation representation and assimilation mechanism. Local, dense-gain, and nonlinear variants share a common framework. Parameter-matched controls, before-first and pre-assimilation prediction, target suppression, subsystem holdouts, and information-destruction tests examine the mechanism.

2. Is dense-gain the preferred realization?

In many evaluated settings, yes. Dense-gain is a simpler and frequently more effective default. In Appendix B.7 (Table 15), it achieves the lowest CRPS among the three standard NBSF variants in all four core belief-quality settings:

SettingDense-gainNonlinear
P12 before-first0.56980.5707
P12 pre-assimilation0.34970.3510
P19 before-first0.55010.5549
P19 pre-assimilation0.33480.3465

Dense-gain is also best in 7/10 subsystem-holdout settings (B.14, Table 25), and is more consistent under target suppression (B.12, Table 22). Hence where information enters the filter can matter more than update complexity.

However, nonlinear assimilation remains useful in some tasks: on P12 mortality it yields AUROC/AUPRC 0.8467/0.5160 versus 0.8312/0.4906 for dense-gain (Table 12), and on patient-grouped eICU it yields 0.8844/0.5791 versus 0.8788/0.5661 (Appendix B.3, Table 11). Thus, dense-gain is our preferred simple default for the tested learned-projection settings, not a universally optimal variant.

3. Are hybrid propagation and latent dimensions required?

No; the ablations do not establish that requirement. On P19, SSM-only propagation achieves AUROC/AUPRC 0.9312/0.6317, compared with 0.9211/0.6117 for hybrid propagation. On P12, their AUROCs are almost identical (0.8465 versus 0.8467; Appendix B.5, Table 13). Similarly, removing the latent block (L=0L=0) increases P19 AUROC from 0.9211 to 0.9249 (Appendix B.6, Table 14). We therefore treat the hybrid residual and latent block as optional capacity choices, not prerequisites for the claimed observation-assimilation mechanism. The results do not justify presenting the complete nonlinear architecture as intrinsically superior.

4. New control: does the update alone explain the gains?

To probe a possible confound, we additionally ran a five-fold P19 fixed-projection ablation. The standard system learns the observation projection WobsW_{\mathrm{obs}} jointly with the update. We replaced it with a fixed coordinate-selection projection, Wobs=[ID∣0]W_{\mathrm{obs}}=[I_D\mid 0]. Mean AUROCs were 0.9255 (local), 0.9241 (dense-gain), and 0.9005 (nonlinear). Under this control, neither cross-variable operator improves mean AUROC over local-coordinate.

This negative control prevents us from attributing learned-projection gains to the update operator in isolation. Our mechanism experiments support the narrower conclusion that jointly learned observation representation and assimilation can exploit cross-variable information, especially with sparse target observations. We will report the full fixed-projection protocol in the camera-ready version if accepted.

5. Clarification of the paper's contribution

We therefore distinguish (i) the probabilistic continuous-discrete belief representation and innovation-driven update, (ii) the empirical role of observation-time information transfer under missingness, and (iii) optional architectural choices. The first two are central to NBSF; the third is task-dependent. Dense-gain is often the more parsimonious implementation, while nonlinear assimilation, hybrid propagation, and latent capacity are alternatives whose benefits must be demonstrated rather than presumed. We will make this hierarchy explicit in the camera-ready presentation if accepted.

Dear Reviewer,

We would be grateful if you could kindly revisit our earlier author response, where we addressed your important questions regarding the essential contribution of NBSF and the role of its architectural components.

Your observations have helped us clarify that the central contribution lies in event-time cross-variable information transfer, rather than the complexity of the nonlinear update.

We have also completed several additional controlled experiments to further substantiate this interpretation. We respectfully present these findings in the following comment for your consideration.

Following our earlier response, we conducted additional controlled experiments to strengthen the evidence for event-time cross-variable assimilation and establish whether the simpler dense-gain update is the preferred realization.

1. Essential contribution and preferred realization

Our central contribution is event-time cross-variable assimilation within a predictive belief-state filter. We agree with your interpretation that the simpler dense-gain update is the preferred realization in settings where this mechanism is beneficial. The nonlinear update, latent block, and hybrid propagation are optional architectural choices, consistent with Appendices B.5–B.6.

2. Fixed observation projection and robustness to missing observations

To isolate the assimilation mechanism from a learned observation projection, we retrained the models with

Wobs=[ID∣0],W_{\mathrm{obs}}=[I_D\mid 0],

verified frozen in every checkpoint.

We evaluated test-time removal of observations across all variables in the first 25% of each stay, using identical masks across models.

Dense-gain minus local AUROC (paired 95% CI; favorable folds):

DatasetNo removal90% removal
P12+0.0148 [+0.0080, +0.0217], 5/5+0.0787 [+0.0681, +0.0891], 5/5
MIMIC-III, ICU stays ≥48\geq48 h−0.0005 [−0.0050, +0.0039], 3/5+0.0440 [+0.0378, +0.0507], 5/5
P19−0.0014 [−0.0069, +0.0041], 2/5+0.0037 [−0.0014, +0.0090], 4/5

The advantage increases substantially under early-observation removal on P12 and MIMIC-III. The MIMIC-III eligible-stay results supersede our earlier all-stays figures. P19 shows no clear improvement.

3. Parameter-matched and multi-seed experiments

We matched the local-coordinate model's capacity to dense-gain (122,684 versus 122,671 parameters).

On P12 at 90% removal, dense-gain exceeds the parameter-matched wide-local model by

\quad 95\%\ \mathrm{CI}=[+0.0994,+0.1227].$$ The advantage persists across three training seeds: | Seed | Dense-gain − local AUROC, 90% removal | |---|---| | 42 | **+0.0787** [+0.0681, +0.0891] | | 43 | **+0.0547** [+0.0450, +0.0643] | | 44 | **+0.0570** [+0.0471, +0.0670] | All five folds favor dense-gain under each seed. These experiments use the same patient folds with different training initializations. Extending the originally capped training runs by up to 200 additional epochs leaves the improvement nearly unchanged. **4. Additional evidence for cross-variable information transfer** We specified a variable-permutation intervention and its success criterion before execution. Observed variable values were independently shuffled across test cells while preserving timestamps and observation masks. At 90% removal, the dense-gain advantage changes from: - **P12:** +0.0787 to −0.0605, a reduction of **0.1392** [0.1244, 0.1535]. - **MIMIC-III:** +0.0440 to −0.0354, a reduction of **0.0794** [0.0705, 0.0885]. Both reductions are observed across all five folds. This provides further evidence that the model uses information in relationships among variables, although the permutation also introduces distribution shift. **5. Full-covariance Kalman comparison** We additionally implemented a full-covariance Kalman-style update with the same propagation network and fixed observation projection. On P12 at 90% removal: - **Dense-gain versus local:** +0.0787 [+0.0681, +0.0891]. - **Full covariance versus local:** +0.0135 [+0.0036, +0.0237]. Full covariance requires approximately **2.0 times the training time and 8.4 times the peak GPU memory** at batch size 64. Although it achieves better one-step-ahead CRPS, it does not reproduce the magnitude of the dense-gain classification improvement. **6. Further clarification of the architecture** On P12 without removal, learned-projection dense-gain and nonlinear achieve similar AUROC (0.8438 and 0.8467). The dense-gain result supersedes the previously quoted 0.8312, which used a different variance-update implementation. The earlier eICU all-stays figures are likewise superseded: on unit stays of at least 48 hours, local-coordinate achieves 0.776, the cross-variable variant 0.771, and dense-gain 0.765 AUROC. These additional experiments reinforce the conclusion that the principal contribution is **where and when cross-variable information enters the filtering process**. Dense-gain provides a simpler realization of this mechanism, with robust classification improvements under missing early observations on P12 and MIMIC-III. We sincerely thank you for your thoughtful feedback and careful evaluation. We hope you will consider our earlier response together with these additional experimental findings when making your final assessment.

Thank you to the authors for their detailed responses and additional experiments. I appreciate their efforts to clarify the central contribution of NBSF and to provide further evidence supporting event-time cross-variable assimilation, particularly under substantial observation missingness.

The new results help address some of my earlier concerns and strengthen the empirical support for the proposed mechanism. However, they also reinforce my original assessment that the main contribution lies in cross-variable information transfer rather than the specific architectural components of the full NBSF model. Moreover, the improvements remain dependent on the dataset and experimental setting, with less consistent benefits on P19 and eICU.

While I acknowledge the value of these clarifications and additional experiments, I do not find them sufficient to change my overall assessment.

Reviewer_J91aOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes a new continuous-discrete filtering method for irregular multivariate time series.

Critical strengths
  1. The paper studies an important and practically relevant problem of modeling irregular multivariate time series. The proposed method is also reasonable.
Critical weaknesses
  1. The paper lacks sufficient comparisons with closely related probabilistic filtering methods, which makes the paper's positioning relative to existing continuous-discrete filtering methods unclear.
Questions for authors
  1. The novelty of the proposed method needs further clarification. Related methods already supports cross-variable updates. The authors should better clarify the methodological advantages of the proposed method.
  2. The experimental evaluation lacks direct comparisons with closely related probabilistic filtering methods, making it difficult to assess the effectiveness of the proposed method.
  3. The writing could be improved, particularly by reducing the use of “;” to connect independent sentences.
3 comments

We thank the reviewer for recognizing the relevance and soundness of NBSF. We clarify its position relative to probabilistic filters.

W1/Q1. Novelty relative to probabilistic continuous-discrete filters

We do not claim that continuous-discrete filtering or cross-variable Bayesian updates are new. A multivariate Kalman filter can update an unobserved coordinate when a correlated variable is measured. GRU-ODE-Bayes, CRU, NCDSSM, and ACSSM also provide related probabilistic approaches.

NBSF makes a more specific contribution. It maintains a factorized Gaussian predictive belief with O(S)O(S) marginal-state storage, propagates it across irregular gaps using a structured mean update and an exact coordinate-wise Ornstein--Uhlenbeck (OU) variance transition, and assimilates measurements through a learned, innovation-driven event-time correction. Unlike Kalman gains, its learned updates do not require propagating an S×SS\times S filtering covariance or computing an exact posterior.

Crucially, we study three controlled assimilation mechanisms with common propagation: local-coordinate, dense linear gain, and nonlinear cross-variable correction. The local control restricts direct updates to the observed coordinate, while the other two permit immediate revision of unobserved belief dimensions. Parameter-matched wide-local controls, before-first and pre-assimilation CRPS, target suppression, and cross-variable information destruction investigate when this event-time transfer is useful, rather than assuming nonlinear coupling is always beneficial. The results favor the simpler dense-gain variant in several settings. Thus, we claim neither universal superiority of the nonlinear update nor isolation of the update operator from the learned observation projection. Our contribution is a lightweight filtering construction and controlled evidence about information transfer in the jointly learned system.

We have clarified these distinctions in Section 2.2 and Section 3.3, including the fact that NBSF's observation updates are filtering-inspired rather than exact Bayesian inference. Figure 4 (Appendix A.4) illustrates the three update variants.

W1/Q2. Direct probabilistic filtering comparisons

To address the request for a closer probabilistic baseline, we additionally evaluated GRU-ODE-Bayes using the authors' implementation on five patient-grouped MIMIC-III folds. The pipelines used the same cached patient cohort and labels, maximum sequence length (300), and identical held-out patient IDs.

MethodAUROC (mean ±\pm SD)AUPRC (mean ±\pm SD)
GRU-ODE-Bayes0.8235 ±\pm 0.00840.4642 ±\pm 0.0186
NBSF local-coordinate0.9040 ±\pm 0.00650.6670 ±\pm 0.0259
NBSF dense-gain0.9021 ±\pm 0.00530.6659 ±\pm 0.0231
NBSF nonlinear0.8995 ±\pm 0.00530.6582 ±\pm 0.0240

All three NBSF variants have higher five-fold mean AUROC and AUPRC. We qualify this comparison carefully: GRU-ODE-Bayes was run with a 20-epoch cap, which four of five folds reached without early stopping. Only fold 5 stopped at epoch 18. Consequently, this is a useful matched-cohort implementation comparison, not proof of superiority over a fully converged or exhaustively tuned GRU-ODE-Bayes model.

We also added a five-fold full-covariance Kalman-style control to test the trade-off against covariance-based assimilation. Under the matched pre-assimilation CRPS protocol, full covariance improves on dense-gain on all five folds of each dataset: P12 0.3214 versus 0.3497, and P19 0.3258 versus 0.3348 (lower is better). Its AUROC is 0.8516 on P12 and 0.9302 on P19. These findings do not support a claim that diagonal filtering is intrinsically more accurate. Instead, they clarify NBSF's design trade-off: it avoids storing and propagating a full covariance, while learning observation-driven cross-variable corrections. This is an internal control, not a direct CRU, NCDSSM, or ACSSM rerun.

We will include the complete new baseline protocols, fold-level results, and convergence qualifications in the camera-ready revision if accepted. The revised manuscript already discusses GRU-ODE-Bayes, CRU, NCDSSM, and ACSSM in Section 2.2 and distinguishes their inference constructions from NBSF.

Q3. Writing and presentation

We agree that the presentation needed clearer, shorter explanations. We revised the description of belief propagation, innovation, and assimilation, clarified the distinction between exact OU variance propagation and learned updates, and added Figure 4 (Appendix A.4) to explain the three variants step by step. We will also further reduce semicolon-linked sentences and shorten dense passages in the camera-ready manuscript. These changes clarify the design and scope of novelty.

Reviewer_J91aOpenReview ↗

Response to authors

I thank the authors for the clarifications and additional experiments, which partially address my concerns. I will keep my current rating.

Dear Reviewer,

Thank you for your thoughtful feedback and constructive suggestions. In response to your questions about NBSF's novelty and relationship to probabilistic filtering methods, we conducted additional experiments and present the results below.

1. Novelty relative to existing probabilistic filters

We agree that cross-variable updates are established in Kalman filtering and related continuous-discrete models. Our contribution is learned innovation-driven event-time assimilation within a factorized Gaussian predictive belief, without propagating a full filtering covariance.

NBSF separates continuous-time propagation from observation-time correction. Its learned assimilation is filtering-inspired rather than an exact Bayesian posterior update.

To evaluate this distinction, we implemented a full-covariance Kalman-style update with the same propagation network and a fixed observation projection:

Wobs=[ID∣0].W_{\mathrm{obs}}=[I_D\mid 0].

The projection was verified frozen in every checkpoint, isolating the assimilation mechanism.

2. Direct comparison with full-covariance Kalman assimilation

We evaluated all three updates on identical P12 folds. Early-observation removal is performed at test time across all variables in the first 25% of each stay, using identical masks.

P12, fixed projectionLocalDense-gainFull covariance
AUROC, no removal0.83620.85100.8397
AUROC, 90% removal0.76540.84330.7813
Before-first CRPS ↓0.61510.58210.6162
One-step-ahead CRPS ↓0.34640.35030.3434

At 90% removal, paired AUROC improvements relative to local are:

  • Dense-gain: +0.0779 [95% CI +0.0675, +0.0884], 5/5 folds.
  • Full covariance: +0.0158 [95% CI +0.0060, +0.0262], 4/5 folds.

Dense-gain provides a substantially larger classification improvement under early-observation removal, persisting across two additional training seeds. It also achieves lower before-first CRPS, whereas full covariance achieves lower one-step-ahead CRPS.

Computational comparison (P12, batch 64):

ModelTraining timePeak GPU memory
Local56.70 s/epoch166 MiB
Dense-gain58.05 s/epoch170 MiB
Full covariance113.25 s/epoch1,421 MiB

Full covariance requires approximately 2.0 times the training time and 8.4 times the peak memory of dense-gain.

3. Comparisons with established probabilistic filtering methods

We evaluated GRU-ODE-Bayes using the authors' implementation on five patient-grouped MIMIC-IV folds with matched held-out stays and labels.

For ICU stays of at least 48 hours:

  • NBSF local-coordinate: AUROC 0.8503 ± 0.0056.
  • GRU-ODE-Bayes: AUROC 0.7519 ± 0.0259.
  • Paired difference: +0.0984 [95% CI +0.0921, +0.1048], 5/5 folds.

This eligible-stay MIMIC-IV comparison supersedes our earlier MIMIC-III comparison, which included shorter stays. GRU-ODE-Bayes used a fixed compute budget rather than exhaustive tuning.

We additionally evaluated NBSF on all 50 Physiome-ODE datasets (250 folds) against published probabilistic-model results:

Published baselineFavourableOverlappingUnfavourable
CRU26240
GRU-ODE4730

These use our prespecified mean ±1 SD interval-separation criterion. They are descriptive comparisons with published results, not equivalence tests or independent baseline reruns.

4. Clarification of predictive-belief comparisons and presentation

We reconciled the CRPS scoring conventions in our earlier response. Full-covariance CRPS of 0.3214 was previously compared with dense-gain CRPS of 0.3497 from an older scorer. Under the matched convention, dense-gain achieves 0.3459; full covariance remains better.

The fixed-projection experiments provide the clearer controlled comparison because forecasting and assimilation share the same coordinate system.

We also clarified the presentation around belief propagation, innovation computation, and event-time assimilation, illustrated in Appendix Figure 4, using shorter and more direct explanations.

We sincerely thank you for your constructive feedback and careful consideration. We hope these additional experiments address your concerns and would be grateful if you would consider them in your final assessment.

Reviewer_cm3vOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper tackles the problem of forecasting in irregularly sampled timeseries. To this end, the authors introduce Neural Belief-State Filtering (NBSF), a continuous-discrete filter that maintains a probabilistic belief over observed and latent variables. The key idea is based on the observation that observation of one variable can directly revise the belief about other unobserved variables at event time. NBSF is deployed in three assimilation variants across four benchmarks, where it is demonstrated to show benifits.

Critical strengths

The paper introduces a novel filtering mechanism that is demonstrated to be beneficial in the forecasting task in irregularly sampled time series.

Experiments performed across multiple real-world benchmarks demonstrate the efficacy of the proposed approach.

Critical weaknesses

The paper is difficult to follow. Even after going through it multiple times, I was unable to understand any intuition behind the design choice of the framework. The sentences are often too long, making it difficult to comprehend the meaning in some cases. The authors should consider rewriting several parts of the paper to make it readable.

It misses several relevant references. I list down some of them below: Klötergens, C. et al. Physiome-ODE: A Benchmark for Irregularly Sampled Multivariate Time-Series Forecasting Based on Biological ODEs. ICLR 2025

Xiao, J. et al. (2024, March). IVP-VAE: Modeling EHR time series with initial value problem solvers. AAAI 2024

Biloš, M. et al. (2023, July). Modeling temporal data as continuous functions with stochastic process diffusion. ICML 2023.

You et al. (2026). Latent Laplace Diffusion for Irregular Multivariate Time Series. ICML 2026

No comparison with the baselines in terms of computation time. Comparison only in terms of peak GPU memory is provided in A.5. It is unclear why OOM is encountered in case of Warpformer even though the peak GPU memory is 8.6 GB

Questions for authors

Why use MIMIC-III instead of MIMIC-IV, which is a newer benchmark?

In the classification task (Table 9), GARLIC outperforms NBSF in 3 of the 4 datasets. How do you then justify the use of NBSF over GARLIC?

Perhaps an illustrative figure could be added to present the key design principle. Figure 2 does not provide any additional information in its current form.

5 comments

We thank the reviewer for the assessment. We address each weakness and question with revised material and additional evidence.

Critical weaknesses

W1. Motivation and readability. NBSF separates belief evolution between observations from belief correction at observation events. For example, unexpectedly low urine output may inform the predicted distribution of BUN even when BUN is unmeasured. The mean propagates through exponential decay and a neural residual; marginal variance follows an exact coordinate-wise Ornstein–Uhlenbeck (OU) transition. When a measurement arrives, its innovation (observed minus predicted value) corrects the belief. The local variant restricts event-time corrections to one coordinate, whereas dense-gain and nonlinear variants can update others. This is known in Bayesian filtering; NBSF studies learned corrections with factorized belief. Figure 4 (Appendix A.4) now illustrates the pipeline and all three variants, complementing Figure 2. Related experiments appear in B.7, B.11, and B.12.

W2. Missing references. We appreciate the suggestions: Physiome-ODE (ICLR 2025) provides biological-ODE forecasting benchmarks; IVP-VAE (AAAI 2024) uses initial-value-problem inference for EHRs; Stochastic Process Diffusion (ICML 2023) and Latent Laplace Diffusion (ICML 2026) study generative temporal modeling. These use different tasks and protocols. We will discuss all four in the camera-ready version if accepted.

W3. Runtime and Warpformer OOM. Appendix A.6 reports NBSF profiling at D=100D=100. Local/dense/nonlinear variants have 198.9/186.6/213.0 ms latency, 151.9/153.7/166.2 MB peak memory, and 198,687/241,903/319,559 parameters. Exploratory cross-model times are 6.34 s/epoch (mTAND), 52.1 (GARLIC), and 292 (NBSF dense-gain). GPU contention and unequal workloads preclude controlled speed claims; NBSF is slower per epoch here. Warpformer's reported 8.6 GB is for P12, whereas its OOM result is for MIMIC-III. We clarified the different settings in Appendix A.6/Table 8.

Questions

Q1. Why MIMIC-III instead of MIMIC-IV? MIMIC-III permits comparison with established published benchmarks. We also completed five-fold patient-grouped MIMIC-IV mortality evaluation, obtaining AUROC 0.9149 (local), 0.9120 (dense), and 0.9096 (nonlinear); local AUPRC is 0.6922. Patient-grouped MIMIC-III AUROCs are 0.9040/0.9021/0.8995 (Table 34), although results are not paired across cohorts. The appendix already evaluates eICU (195,495 records) with five-fold patient-grouped AUROC/AUPRC 0.8829/0.5733 (local), 0.8788/0.5661 (dense), 0.8844/0.5791 (nonlinear) (B.3/Table 11), plus hospital-grouped nonlinear results 0.8804/0.5565. These extend evaluation beyond MIMIC-III. The complete MIMIC-IV protocol and folds will be included in the camera-ready paper if accepted.

Q2. Why NBSF relative to GARLIC? The comparison comprises three shared clinical tasks, not four datasets. Reported AUROC/AUPRC (%, Tables 3 and 9) are:

TaskGARLICNBSF nonlinearNBSF dense
P12 mortality86.40/56.8984.67/51.6083.12/49.06
P19 sepsis90.96/55.2992.11/61.1793.00/62.39
MIMIC-III mortality90.09/64.8589.95/65.8290.21/66.59

Dense-gain has higher reported P19 values by 2.04 AUROC and 7.10 AUPRC points, and higher MIMIC-III values on both metrics; GARLIC leads on P12 mortality. However, P19/MIMIC-III cohort and feature definitions differ; these are published-reference, not matched-fold, comparisons.

More importantly, GARLIC learns relational representations for prediction, while NBSF maintains an explicit Gaussian predictive belief and uses innovations to revise it at measurement times. Our contribution includes evaluating when this information transfer is useful: dense-gain lowers pre-assimilation CRPS versus local on P12 (0.3497 vs 0.3753) and P19 (0.3348 vs 0.3430) (B.7/Table 15). On patient-grouped MIMIC-III, nonlinear before-first CRPS improves in all five folds, with bootstrap intervals excluding zero (B.11/Table 21); target suppression is evaluated in B.12/Tables 22–23. Beyond ICU cohorts, PAMAP2 eight-subject leave-one-subject-out experiments show before-first CRPS decreasing 0.5477 to 0.5411 at 10% missingness and 0.5382 to 0.5204 at 40%, with all eight subjects favoring cross-variable in each setting (Wilcoxon p=0.0078p=0.0078; B.15/Tables 26–27). This supports mechanism-level, not universal superiority, claims.

Q3. Illustrative figure. We added Figure 4 in Appendix A.4, which depicts irregular measurements, mean/variance propagation, masked innovation, reliability weighting, and all three event-time updates. The equations distinguish exact OU variance propagation from learned (not exact Bayesian) assimilation.

Reviewer_cm3vOpenReview ↗

After rebuttal

I acknowledge the rebuttal from the authors, but they don't generally address my concerns. I don't agree that the reason for using MIMIC-III instead of MIMIC-IV is just to make a comparison with the existing work easier. Simply implement these baselines on MIMIC-IV. Also, the related papers mentioned in the review are supposed to be used as baselines rather than discussed in the related work. I also don't think minor changes to writing solve the readability issues in the paper. So, I will stick with my current rating.

Dear Reviewer, Thank you for your detailed review and subsequent comments. In response to the request for additional MIMIC-IV baselines and a stronger comparison with GARLIC, we conducted new matched experiments. The results are presented below.

1. Direct comparison with neural baselines on MIMIC-IV

We evaluated GARLIC, mTAND, GRU-D, and GRU-ODE-Bayes alongside NBSF using the same five patient-grouped folds, matched held-out stays, and identical outcome labels.

The evaluation is restricted to the recorded eligible ICU stays corresponding to the 48-hour observation window. These results supersede our earlier all-stays figures, which included shorter stays with observations extending beyond ICU discharge.

Table 1. MIMIC-IV mortality prediction (five-fold mean ± SD).

ModelAUROCAUPRC
NBSF local-coordinate0.8503 ± 0.00560.5245 ± 0.0060
NBSF dense-gain0.8466 ± 0.00580.5160 ± 0.0126
GARLIC0.8442 ± 0.00620.5125 ± 0.0046
mTAND0.8434 ± 0.00350.5099 ± 0.0078
GRU-D0.8382 ± 0.00400.5036 ± 0.0158
NBSF nonlinear0.8381 ± 0.00440.4944 ± 0.0089
GRU-ODE-Bayes0.7519 ± 0.02590.3776 ± 0.0304

NBSF local-coordinate achieves the highest AUROC and AUPRC among the evaluated neural models. Matched patient-clustered bootstrap comparisons further quantify its performance relative to established neural baselines:

ComparisonΔ\DeltaAUROC [95% CI]Folds favouring NBSF
NBSF local − GARLIC+0.0061 [+0.0024, +0.0100]4/5
NBSF local − mTAND+0.0069 [+0.0033, +0.0105]5/5
NBSF local − GRU-D+0.0121 [+0.0084, +0.0159]5/5

All models were trained on their original training folds and evaluated on the common eligible test subset. Model selection used validation data. GRU-ODE-Bayes was evaluated under a fixed computational budget rather than exhaustive hyperparameter optimization.

2. Why use NBSF rather than GARLIC?

We additionally reran GARLIC using its published P12 configuration on our patient folds.

GARLIC achieves an AUROC of 0.8471±0.00690.8471\pm0.0069. Fixed-projection NBSF dense-gain achieves a paired AUROC difference of

\qquad 95\%\ \mathrm{CI}=[-0.0026,+0.0101].$$ Thus, the two implementations show no clear discrimination difference on P12. On MIMIC-IV, NBSF local-coordinate exceeds GARLIC with a paired confidence interval excluding zero. Beyond classification performance, NBSF provides an explicit predictive belief over state variables and supports incremental assimilation at observation times. In our streaming evaluation, the CPU latency from a new observation to an updated prediction is: - **NBSF:** $0.33$ ms. - **GARLIC:** $25.71$ ms. This represents an approximately **78-fold latency difference** in the evaluated implementations. GARLIC's available implementation re-encodes the observation history, whereas NBSF updates a cached belief state. These experiments establish a direct predictive comparison and a computational distinction between the methods. The following comment presents our additional results on the suggested forecasting literature, controlled computational benchmarks, and the methodological intuition underlying NBSF.

Comments on the paper

1 comment

Dear Area Chair,

We have carefully addressed the concerns and questions raised by the reviewers through detailed rebuttal responses, methodological clarifications, and additional experiments.

We would be grateful if you could kindly encourage the reviewers to consider our responses and the additional experimental evidence before finalizing their assessments.

Thank you for your valuable time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.