PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Zhen Xu, Nicholas Konz, Zhen Tan, Zachary Plotkin, Tianlong Chen

arXiv:2610.11090v1cs.CE

TL;DR

ICU models can learn dose–outcome relationships that reflect clinicians’ responses to severity rather than drug effects, limiting prediction under changed dosing practices. PhysioWorld isolates the dose contribution with a paired forward pass and projects it onto pharmacological signs; it achieves the lowest out-of-distribution RMSE across evaluated settings.

  • Problem

    Because clinicians titrate treatment to patient state, ICU models may learn dose–outcome associations that do not represent the drug’s effect and generalize poorly to changed dosing practices.

    Source: §1 · pp. 1–2

  • Method

    PhysioWorld subtracts a zero-dose forward pass from the recorded-dose pass and projects the resulting intervention effect onto pharmacologically specified signs.

    Source: §1 · p. 2, §4 · pp. 4–5

  • Results

    Across five distribution-shift axes and three ICU databases, PhysioWorld exceeds every baseline out of distribution, reducing RMSE by up to 12% over the strongest external baseline.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The reported generalization gain depends on pharmacological knowledge: inverted or data-mined directions perform worse than using no corpus.

    Source: §5.5 · p. 9, §6 · p. 10

    The projection constrains signs but not magnitudes, and its guarantee applies to G∆z rather than decoded predictions.

    Source: §4 · p. 6

Abstract

from arXiv · show

Predicting how a patient's vital signs 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. However, such models tend to be brittle outside of 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, generalizing poorly to out-of-distribution (OOD) dosing practices. Yet, pharmacological information provides the directional effects of different drugs, which are invariant to OOD shifts. To resolve the OOD sensitivity of current ICU world models, we introduce PHYSIO WORLD, which isolates an intervention's effect by evaluating a forward pass twice: once under recorded doses, and once with those doses set to zero. The difference is projected onto pharmacologically admissible effect directions stated by 70 rules over 28 drugs, leaving the magnitude to be learned from data. Across five distribution-shift parameters over cohorts from three intensive-care databases, and against forecasters, counterfactual models, and invariance objectives, PHYSIO WORLD attains the lowest OOD RMSE in every setting, improving on the strongest external baseline by 8-12%, while matching the in-distribution error of its backbone. Reversing the pharmacological directions degrades accuracy below the backbone, and data-mined directions recover almost no gain, indicating the improvement derives from the content of the pharmacological knowledge. The construction may extend to other settings where the sign of an effect is known in advance and its magnitude is confounded by treatment assignment.

1 Introduction

ICU world models must predict physiology under specified treatment courses, but observational dose–outcome associations can reflect patient severity rather than drug effects. PhysioWorld constrains intervention effects using pharmacological directions, improving out-of-distribution prediction across dosing and patient shifts.

1 Introduction

  • Because clinicians titrate doses to patient state, models can learn collapsed or reversed drug effects that survive only the training regime.A vasopressor may be associated with falling pressure because it is given as the patient deteriorates.Source: §1 · pp. 1–2
  • Models degrade more under treatment-intensity shifts than under shifts in patient population, while PhysioWorld exceeds every baseline out of distribution at its backbone’s in-distribution error.The gain holds across five shift axes and three ICU databases.Source: §1 · p. 2
  • Pharmacological effect directions remain invariant across dosing shifts, even though effect magnitudes vary with age, organ function, and comorbidity.The paper uses these directions as prior knowledge rather than estimating them from observational dose associations.Source: §1 · p. 2
  • PhysioWorld estimates an intervention’s contribution with a second forward pass at zero dose, then projects the difference onto pharmacologically admissible directions.Its parameter-free projection constrains signs while leaving effect magnitudes to training.Source: §1 · p. 2

2 Related Work

Prior work supplies clinical sequence models, counterfactual estimators, shift-robust objectives, and knowledge constraints, but the paper identifies a gap: these approaches do not constrain the intervention effect inside latent dynamics.

World models

  • Medical world models reviewed here do not support the specified intervention-course task with pharmacological structure and distribution-shift evaluation.Among six, four omit intervention inputs, one predicts an empty outcome, and one uses a binary vasopressor flag.Source: §2 · p. 2

Clinical prediction, interventions and shift

  • Clinical forecasters model interventions as covariates, while counterfactual models adjust for time-dependent confounding and invariant-feature methods target shifts in patient identity, unit, or era.The paper treats these counterfactual models as baselines and notes their exposure to treatment confounding.Source: §2 · pp. 2–3

Knowledge-constrained learning

  • Existing knowledge-constrained methods constrain predicted values or input derivatives, but those constraints can miss a dose effect already propagated through latent representations.The paper therefore places its constraint on the intervention effect within the recurrent transition.Source: §2 · p. 3

3 Problem Setup

The task is to predict the next observed vital-sign state from a patient’s history and a specified future intervention course. The setup represents irregular observations, asynchronous interventions, and distribution shifts explicitly.

Problem formalization

  • Each patient stay contains irregularly timed vital-sign states with freshness masks, asynchronous intervention events, and static covariates.The mask distinguishes newly measured values from values carried forward between measurements.Source: §3 · p. 3
  • The model predicts the next observed state from the history and interventions scheduled between the current and next observation times.In free-running multi-step prediction, predictions feed back as states while the specified future intervention plan remains fixed.Source: §3 · p. 3

Out-of-distribution evaluation

  • Out-of-distribution error is measured by fitting on one cohort slice and scoring held-out stays in other slices along a selected shift axis.The diagonal train/test slice is in-distribution; the off-diagonal entries are the reported out-of-distribution errors.Source: §3 · p. 3

Stream representation

  • Asynchronous interventions are represented as bucketed dose histories across five elapsed-time windows, rather than with a pharmacological elimination kernel.The same coarse timescale applies across channels, avoiding fitted per-drug constants in the input representation.Source: §3 · p. 4

4 PhysioWorld

PhysioWorld separates the dose contribution from the predicted latent transition, then projects that effect onto signs supported by a pharmacological rule corpus. The projection is exact and non-expansive, but its guarantee is limited to signs in the calibrated latent channel space.

4 PhysioWorld

  • The rule corpus specifies effect signs for 70 drug-channel pairs across 28 drugs, omitting magnitudes because observational dose associations are confounded by titration.Physiological value ranges and channel ordering are not used as constraints.Source: §4 · pp. 4–5
  • A second pass with covered drug doses zeroed isolates the dose-history contribution while keeping weights, observations, and non-dose intervention channels fixed.The paired recurrent states diverge after the first administered dose; the extracted effect is dominated by the recent past, with a three-step window recovering it to within 1%.Source: §4 · p. 5
  • The effect difference is mapped into channel units, and only channels with agreeing signs among active covered drugs are constrained.Conflicting signs, such as simultaneous vasopressor and sedative effects, are left unconstrained; conflicts affect 2.3% of scored entries and 4.1% under treatment.Source: §4 · p. 5
  • The projection correction is added to the latent state advanced into the next recurrent step, rather than applied only to the decoded output.The sign projection is the Euclidean projection onto a nonempty closed convex cone and is non-expansive.Source: §4 · pp. 5–6
  • The guarantee fixes signs, not magnitudes, and applies to G∆z rather than decoded predictions; conflicts and corpus-silent effects receive no constraint.The projection moves responses on 54–61% of scored entries where a sign is defined.Source: §4 · p. 6
  • The backbone is trained as a JEPA over represented streams, while the readout is trained after the backbone is frozen and the calibration map is anchored.The projection reads a detached calibration map so the prediction loss cannot rotate it to make the constraint vacuous.Source: §4 · pp. 4, 6

5 Experiments

The experiments test whether pharmacological sign constraints improve ICU world-model generalization across population and treatment shifts, including free-running rollouts and dose perturbations. PhysioWorld performs best across the reported shift axes and retains its advantage in rollout and dose-response evaluations; ablations indicate that both the pharmacological directions and the two-pass projection matter.

5.1 Setup

  • The benchmark uses five cohorts from three databases and five shift axes; it compares forecasters, counterfactual models, and invariance objectives using OOD RMSE and skill against persistence.Axes include age, disease subtype, era, vasopressor intensity, and sedative intensity; cross-database comparison uses sedative shift alone because VitalDB has too few vasopressor-exposed stays.Source: §5.1 · pp. 6–7

5.2 Shift axes: population against intervention intensity

  • PhysioWorld is best on all five shift axes, matches its backbone’s in-distribution error, and exceeds the best baseline on each of five sedative-shift cohorts.Its margin over Clamp is positive on every cohort, but within about two seed standard deviations on four; the reported ordering is consistent, not a win on every individual cohort.Source: Figure 3 · p. 8, §5.2 · p. 8

5.3 Horizon analysis under free-running rollout

  • PhysioWorld has the lowest rollout error at every horizon (8.52 at h=1; 10.28 at h=8) and the flattest learned-model growth (+0.090).Its skill is 0.258 at h=1 and 0.247 at h=8; its dispersion is 0.842, closest to one among learned models.Source: Figure 4 · p. 8, Table 2 · p. 8, §5.3 · p. 9

5.2 Shift axes: population against intervention intensity

  • Population shifts cause moderate degradation, while under intervention-intensity shifts sequence models fall below persistence (PatchTST 12.15, persistence 10.83) and explicit-intervention models remain competitive (Causal Transformer, 10.45).PhysioWorld’s advantage over Clamp is consistent but modest across Age, Disease, and Era.Source: §5.2 · pp. 7–8

5.4 Dose response under perturbation

  • PhysioWorld’s dose-response agrees with the pharmacological direction on 86% of eligible steps under zeroing and 79% under permutation, versus at most 77% and 70% for any baseline.Its sensitivity is comparable to forecasters, so the higher agreement is not obtained by responding less.Source: Table 3 · p. 9, §5.4 · p. 9

5.5 Knowledge and component ablations

  • Removing the sign projection or second forward pass lowers full-method skill from +0.20 to +0.11 and +0.08, respectively, versus +0.05 for the backbone.Withholding rule pairs degrades skill to 0.152 at half the corpus and 0.107 at a quarter.Source: §5.5 · p. 9
  • Skill falls from 0.198 with full direction agreement to −0.058 with all directions reversed, below the backbone’s 0.053; mined directions score 0.005.Mined directions agree with pharmacology on only 48% of the 70 pairs, while magnitude-band rules score 0.093.Source: Table 4 · p. 9, §5.5 · p. 9

6 Conclusion

The conclusion presents PhysioWorld as a way to reduce error when treatment practice differs from training records, by isolating intervention effects and constraining their signs to pharmacological directions.

6 Conclusion

  • PhysioWorld reduces error under treatment-practice shifts, with its largest gain under treatment-intensity shifts, across five axes and three databases.Inverted or data-mined directions perform worse than no corpus, supporting the conclusion that the gain reflects pharmacological content rather than the constraint alone.Source: §6 · p. 10

B Rule Corpus

The rule corpus specifies pharmacological effect signs without magnitudes, while stream representation and channel-level coverage determine when those constraints apply.

B Rule Corpus

  • The corpus contains 70 sign rules across 28 drugs, and its stated coverage per drug ranges from 9.5% to 18.0%.Four drugs intentionally have rules in both directions because their pressure effects and reflex heart-rate effects can oppose one another.Source: B Rule Corpus · p. 15
  • Conflict marks scored entries where active-drug rules disagree; Silent marks entries where no active drug supplies a rule for that channel.Conflict/treated divides the conflict share by entries with at least one active corpus drug, measuring how often treatment leaves the channel unconstrained.Source: B Rule Corpus · p. 15
  • Three channels carry 58 of the 70 rules, while SpO2 has none and each systolic or diastolic pressure channel has only one.Coverage is limited when no rule applies or when active drugs impose opposing directions; propofol with norepinephrine is a common opposing-drug combination.Source: B Rule Corpus · p. 16

C Stream Representation

  • Drug exposure is represented by doses summed within five elapsed-time windows, while infusion rates are integrated over each window rather than counted as impulses.Ventilator settings bypass dose bucketing, and channel freshness is encoded separately with a fixed six-hour time constant.Source: C Stream Representation · p. 16

D.1 Calibration map sensitivity

Calibration sensitivity tests show how the anchor, row orthonormality, and fitting protocol affect the learned map. Additional checks examine how much dose history is needed and which interventions are withdrawn.

D.1 Calibration map sensitivity

  • Axis-initialized rows reached the default solution at median cosine 0.97, with no sign reversals and unchanged OOD error of 8.70.A random orthonormal control reached OOD error 9.24, above the default at 8.69.Source: §D.1 · pp. 17–18, Table 8 · p. 18
  • Weakening the anchor term tenfold raised OOD RMSE to 8.89 and reduced between-seed agreement to 0.71, while strengthening it tenfold changed little.The weakened setting also produced two sign reversals; the default therefore lies on a plateau rather than at a finely tuned point.Source: §D.1 · p. 17
  • Unnormalized rows cost 0.25 RMSE, while detaching G is retained as a precaution against an unruled-out rotation that could make the constraint vacuous.Unnormalized rows preserve binding rate, but the resulting write-back is not minimum-norm and can let constraints reach other channels.Source: §D.1 · pp. 17–18

D.2 How far back the second pass reaches

  • At w = 3, the short window misses 1% of the effect and preserves 99.3% of clipping decisions; slow channels remain a conservative approximation.Figure 5 plots relative state-change error and decision inconsistency across window sizes, separately for in-distribution and shifted slices.Source: Figure 5 · p. 18, §D.2 · p. 18

D.3 What the second pass withdraws

  • The two passes withdraw drug doses alone; removing ventilator settings or status flags would instead ask what happens with ventilation switched off, without rules for those effects.Source: §D.3 · p. 19

D.4 Why the intervention stream stays in the recurrent core

  • The intervention stream stays in the recurrent core because intervention history also carries information about the patient’s state, such as six hours of norepinephrine exposure.Source: §D.4 · p. 19

E Component Ablation

Component ablations show that the full gain depends on both isolating the intervention effect and projecting it onto pharmacological directions; training details specify how the calibration map is maintained.

E Component Ablation

  • The ablation removes one component at a time with the rule corpus fixed, reporting skill and RMSE averaged across all five shift axes.Source: E Component Ablation · p. 19
  • Removing sign projection retained 40% of the gain over the backbone, while removing the second pass left skill within 0.03 of the unconstrained backbone.Neither component alone recovers the full method’s gain.Source: E Component Ablation · p. 19

F Hyperparameters and Training Details

  • The calibration map has orthonormal rows, is trained only by an anchor loss, and is re-orthonormalized after every optimizer step.The projection uses a detached copy of the map; its added weights total 288 parameters, as described in the implementation settings.Source: F Hyperparameters and Training Details · p. 19

G Holdout Construction

Holdout construction evaluates models across population and treatment-intensity slices, with minimum-size and availability rules shaping which comparisons are possible. The accompanying sections report computational cost and convergence checks.

G Holdout Construction

  • Models are trained on one slice and tested across every slice in the same partition, with off-diagonal results treated as out-of-distribution error.Source: G Holdout Construction · p. 19
  • Population slices use sex, age, care unit, diagnosis subtype, and—on MIMIC-IV—era strata.Age follows five bins, and diagnosis subtype is grouped by ICD-9/10 circulatory-system categories.Source: G Holdout Construction · pp. 19–20
  • Treatment-intensity slices rank each drug’s peak dose among exposed stays, use the maximum rank within a class, and compare low and high burden bands.The middle quartile band is a buffer; zero-exposure stays receive no burden score and enter no intensity slice.Source: G Holdout Construction · p. 20
  • Slices below 450 admissions are excluded, training sets are balanced to the smallest slice, and some cohort axes are unavailable or yield too few slices.These availability constraints rule out particular drug-class comparisons on MIMIC-III, MIMIC-IV injury, and VitalDB.Source: G Holdout Construction · p. 21

H Cost

  • PhysioWorld training takes 46.9 ms per step versus 21.0 ms for the JEPA backbone; inference takes 25.2 ms versus 6.6 ms.Chronos-2 inference takes 7,839 ms per step; PhysioWorld adds 288 calibration-map parameters and retains recurrent O(Td^2) complexity.Source: H Cost · p. 22, Table 13 · p. 22

I Convergence

  • Every architecture had reached a validation-loss plateau by the end of its training budget; last-quarter gains were 0.01–0.45% against total gains of 0.9–11.9%.Source: I Convergence · p. 22

J Proof of Proposition 4.1

The proof identifies Π as the Euclidean projection onto the admissible cone and establishes its non-expansiveness and squared-error improvement property.

J Proof of Proposition 4.1

  • Π is the Euclidean projection onto C, acting coordinate-wise as σc max(σct, 0) for constrained coordinates and leaving unconstrained coordinates unrestricted.The constraint set is a Cartesian product of closed convex cones, including R for coordinates without a specified sign.Source: J Proof of Proposition 4.1 · p. 22
  • Each coordinate map is 1-Lipschitz, so summing squared coordinate differences establishes that Π is non-expansive.The coordinate maps are the identity, max(·, 0), or min(·, 0).Source: J Proof of Proposition 4.1 · p. 22
  • Projection reduces squared error to any admissible y⋆ by at least the squared distance from y to Π(y).The decomposition uses the projection theorem to establish a nonnegative cross-term.Source: J Proof of Proposition 4.1 · p. 23

K Significance against Clamp

PhysioWorld is compared with Clamp using patient-level bootstrap intervals and a slice-pair Wilcoxon test. The section also describes the horizons and dispersion reported for autoregressive rollout.

K Significance against Clamp

  • The primary test resamples test patients 1,000 times; the Wilcoxon test covers 56 slice pairs but is approximate because pairs share training slices.The patient-level bootstrap uses the same resample for both methods and does not assume independence across slice pairs.Source: K Significance against Clamp · p. 23
  • On 48 of 56 slice pairs, PhysioWorld has lower error, with a median RMSE difference of 0.11 and one-sided Wilcoxon p < 10−6.Table 14 reports Clamp − PhysioWorld RMSE differences with patient-level 95% bootstrap intervals; positive values favor PhysioWorld.Source: Table 14 · p. 23, K Significance against Clamp · p. 23

L Rollout at every horizon

  • Table 15 reports autoregressive RMSE and across-seed standard deviation at every horizon, plus prediction dispersion at the first and last steps.The rollout follows a half-stay warmup; the slope fits log RMSE against log horizon for h ∈ {1, 2, 4, 8}.Source: L Rollout at every horizon · p. 24, Table 15 · p. 24

Related papers

Loading 2610.11090v1…