acceptodds
Under review as a conference paper at ICLR 2027

Sign-Constrained Intervention Effects for Domain-Generalizable ICU World Models

Abstract

Predicting how a patient's vital signs will respond to an intervention is a central question in intensive care. World models can do so by learning dynamics as a function of prior actions, so that a trajectory can be unrolled under any chosen treatment course. However, such models tend to be brittle for dosing practices outside of those seen in training data. Clinicians choose drug dosages based on the patient's state, so the association a model learns between dose and outcome runs opposite to the drug's effect, which does not generalize well to out-of-distribution (OOD) dosing practices. However, pharmacological information already provides the effects on vital signs of different drugs, which is invariant to dosing practice distribution shifts. To resolve the OOD sensitivity issues of current ICU world models, we introduce **PhysioWorld**, a world model that isolates an intervention's effect by evaluating a forward pass twice: once under the recorded doses, and once with those doses set to zero. The difference is then projected onto the set of pharmacologically admissible *effect directions* stated by a corpus of 70 (drug, channel, sign) rules over 28 drugs, leaving the magnitude, which no source states, to be learned from data. Across five distribution-shift parameters over cohorts from three intensive-care databases, and against forecasters, counterfactual treatment-effect models, invariance objectives and two weaker placements of the same knowledge, PhysioWorld attains the lowest OOD RMSE in every setting, improving on the strongest baseline by 8–12%, while maintaining the in-distribution error of an unconstrained backbone's. Reversing the pharmacologic effect directions degrades accuracy below the unconstrained backbone, and directions estimated from the training records recover almost no performance gain, indicating that the improvement derives from the content of the pharmacologic knowledge rather than from the presence of a constraint. Nothing in the construction is specific to intensive care: it applies wherever the sign of an effect is known in advance and its magnitude is confounded by how the treatment was assigned.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
world models, clinical time series, out-of-distribution generalization, knowledge-constrained learning, action-conditioned latent dynamics

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

The authors develop PhysioWorld to constrain the direction of a model’s predicted response to medication. Experiments across five ICU cohort shifts report improved forecasting and dose-response agreement.

Critical strengths

The main contribution is a simple way to incorporate pharmacological knowledge into a predictive model. Evaluation across several databases and the dose-perturbation analysis provide useful tests of performance under changes in treatment practice.

Critical weaknesses

The ablation leaves the source of improvement unclear. Removing projection gives z(0) + [z(u) − z(0)] = z(u), yet Table 9 attributes 40% of the gain to the two-pass calculation...

Questions for authors

Please match anchor supervision and training settings between JEPA and the projection-removed model, and explain the remaining gain. Also, calify which experiment produced the creatinine curve, and if Figure 5’s diagnostics hold for the stated prediction targets.

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

This paper investigates the out-of-distribution (OOD) generalization of intervention-conditioned ICU world models, focusing on the confounding between treatment intensity and patient severity in observational clinical data. The authors propose PhysioWorld, a JEPA-based recurrent dynamics model that incorporates external pharmacological knowledge to constrain predicted intervention effects. Specifically, the model performs two forward passes using the recorded drug doses and zeroed drug doses, respectively, and interprets the difference between their latent predictions as an estimated intervention effect. A learned calibration matrix maps this difference into physiologically meaningful dimensions, where a sign projection enforces 70 predefined drug–vital-sign directional rules covering 28 interventions. The corrected representation is then used for state prediction and multi-step rollout. Experiments across five patient cohorts from three clinical databases evaluate generalization under population and treatment-intensity shifts. The reported results show improvements in OOD prediction accuracy over existing baselines, improved multi-step forecasting performance, and the importance of incorporating correct pharmacological knowledge.

Critical strengths
  1. The paper introduces a meaningful approach to incorporate pharmacological knowledge into patient dynamics modeling. Instead of directly constraining changes in vital signs between consecutive time steps, the proposed method constrains the estimated intervention effect by comparing predictions with and without treatment. This formulation is conceptually reasonable because physiological changes may result from both disease progression and treatment effects.

  2. The paper provides a comprehensive empirical evaluation of the proposed approach. Beyond standard prediction metrics, the authors investigate the role of pharmacological knowledge through rule ablations and dose perturbation experiments, including removing, reversing, and modifying the injected rules. The evaluation across multiple clinical cohorts and distribution-shift settings further strengthens the empirical evidence for the proposed method.

Critical weaknesses
  1. Although the rule ablations demonstrate the importance of correct pharmacological knowledge, the experiments do not sufficiently isolate the contribution of the proposed mechanism from that of the injected knowledge itself. In particular, the Clamp baseline, which also incorporates pharmacological rules, already achieves substantial improvements over the unconstrained JEPA backbone, while the additional gains of PhysioWorld in single-step prediction are relatively modest. Although the multi-step results provide encouraging evidence, the paper does not adequately establish why latent-space projection is preferable to directly constraining the decoded intervention-effect difference and feeding the corrected predictions back into the dynamics model. Similarly, the necessity of introducing a separate calibration matrix GG, rather than leveraging the existing decoder, is not sufficiently examined. More controlled comparisons using the same pharmacological rules and comparable training settings would help clarify the contributions of these design choices.

  2. The validity of the estimated intervention effects remains insufficiently verified. The proposed method estimates treatment effects through two forward passes under recorded and zeroed drug doses. However, the zero-treatment branch does not have corresponding ground-truth counterfactual supervision, and both branches condition on the same observed patient history, which may already reflect the effects of previous treatments. Consequently, their difference does not necessarily represent the true causal effect of treatment. The current evaluations primarily measure factual prediction accuracy, response sensitivity, and agreement with pharmacological rules, but do not directly assess the accuracy of the zero-treatment predictions or the estimated treatment-effect magnitudes. Although the authors acknowledge this limitation, the claims regarding intervention-effect isolation and what-if prediction would benefit from clearer qualification.

Questions for authors
  1. How much of the performance improvement comes from the pharmacological knowledge itself versus the proposed latent-space mechanism? While the Clamp baseline and component ablations provide partial evidence, it remains unclear whether latent-space projection offers additional benefits over directly projecting the decoded intervention-effect difference and feeding the corrected predictions back into the dynamics model. Could the authors clarify this distinction and provide more controlled comparisons to justify the latent projection and separate calibration matrix GG?

  2. How reliable is the zero-treatment branch as an approximation of counterfactual outcomes? Since the model zeros out drug doses while retaining the observed patient history, which may already reflect previous treatments, the resulting difference may not correspond to the true treatment effect. Could the authors provide additional evidence, such as evaluation in a controlled simulation with known counterfactual outcomes, or clarify the intended interpretation and limitations of the estimated intervention effects?

Reviewer_4wfnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PhysioWorld, an ICU world model that predicts next-step vital signs under a given intervention course. The main argument is that dose is confounded by indication, so data-driven models learn a wrong or empty dose-response and break under shifts in treatment intensity. PhysioWorld runs a second forward pass with the doses zeroed, takes the latent difference as the intervention effect, and projects it onto the signs given by a rule corpus of 70 (drug, channel, sign) rules over 28 drugs. Magnitudes are still learned from data. On five cohorts from MIMIC-IV, MIMIC-III, and VitalDB, it gets the lowest OOD RMSE on all reported shift axes, and the corpus ablations suggest the gain comes from the pharmacological content.

Critical strengths

● The idea is simple and well motivated. Confounding by indication is a real problem in ICU data, and using only the sign of a drug's effect is a sensible choice of prior knowledge. The projection has no parameters, comes with clear guarantees, and is cheap. ● The evaluation is careful: three databases, five seeds, balanced training sets, and patient-level bootstrap tests. The corpus ablation in Table 4 is the most convincing result. Skill drops steadily as more rules are inverted and goes below the backbone when all are inverted, so the knowledge itself matters.

Critical weaknesses

● I could not tell where the gain comes from. Without the sign projection, equation 3 reduces to ẑ(u), which is just the backbone. Yet "−sign projection" in Table 9 improves skill from +0.05 to +0.11. The only difference I can see is the anchor term on G. If that term is not used for JEPA, Soft constraint, and Clamp, the comparison is not on the same backbone. Also, "−second forward pass" (RMSE 10.08) and Clamp (about 8.84 averaged from Table 1) both act on the output but differ a lot. Finally, the rules define a sign on only 13% of scored entries, yet OOD skill almost quadruples. I would like to see errors on covered vs. uncovered entries. ● The paper motivates the model with what-if queries, but every result uses the recorded doses, so counterfactual accuracy is never tested. The agreement metric in Table 3 checks agreement with the injected rules, which is close to circular. The closest knowledge-based methods the paper cites (Hussain et al., 2021; Qian et al., 2021) are not compared. Against Clamp, the bootstrap interval excludes zero on only 4 of 9 cohort-axis pairs, and on none outside MIMIC-IV circulatory. The Sex, Care unit, Rate control, and Vasodilator axes are built in Appendix G but not reported.

Questions for authors

● What exactly differs between "−sign projection" and the JEPA backbone in Table 9? Is the anchor term responsible for the gain, and do the same-backbone baselines use it? ● How is Clamp defined, and why does it differ so much from "−second forward pass"? ● Can you report OOD error separately for entries where a sign is defined and where none is? ● Why are the Sex, Care unit, Rate control, and Vasodilator axes not reported?

1 comment

Note: No rebuttal posted yet.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.