PhysioWorld

TL;DR Clinicians dose against severity, so ICU models learn drug effects backwards and break when dosing practice shifts. PhysioWorld isolates each intervention's effect with a second forward pass and constrains only its sign to pharmacology, lowering out-of-distribution error by 8–12% over the strongest external baseline.

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 external baseline by 8–12%, while matching the in-distribution error of an unconstrained backbone. Reversing the pharmacological 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 pharmacological knowledge rather than from the presence of a constraint.

Why ICU world models fail under dosing shift

A patient stay is two streams on one clock: irregularly charted vital signs and sparse clinician actions. A world model must predict the next vitals under a treatment course one names.

But each recorded dose is itself a reading of how sick the patient was. A vasopressor appears because blood pressure was falling, so a model that reads the dose as a cause learns that the drug makes pressure fall. Such a dose–response survives only the dosing regime it was fitted in: under a shift in treatment intensity, pure sequence models fall below simple carry-forward persistence.

The direction of a drug's effect, by contrast, is invariant across that shift: a vasopressor raises blood pressure across patients and care units, and pharmacology states as much before any data is collected.

A real ICU stay: vital-sign traces over time with drug administrations marked below. — A real ICU stay illustrating the task.

Method

PhysioWorld separates what a forward pass predicts from what pharmacology is able to constrain. The rules, the sign cone they define and the projection onto it are fixed before training; the backbone and the calibration map G that reads effects in channel units are learned.

STEP 1

Rule corpus

70 (drug, channel, sign) rules over 28 drugs state the direction in which each drug moves each vital sign. Magnitudes are left out, because dose is titrated against the very quantity it changes.

STEP 2

A second forward pass

The same predictor, with the same weights and history, runs with the recorded doses and with those doses zeroed. Their difference Δz = ẑ(u) − ẑ(0) is the model's own claim about what the treatment contributed.

STEP 3

Sign projection

A closed-form, parameter-free projection clips any component of GΔz that contradicts the stated sign and writes the correction back into the latent, inside the recurrence. It is provably non-expansive.

Results

8–12%

lower OOD RMSE than the strongest external baseline

5 / 5

shift axes on which PhysioWorld is best

86%

dose responses in the pharmacological direction (best baseline 77%)

−0.058

skill with every rule inverted, below no rules at all (0.053)

Population shift versus intervention-intensity shift

Out-of-distribution RMSE on MIMIC-IV (lower is better, mean over five seeds). Underlined: best external baseline on that axis.

	Population shift			Intervention-intensity shift
Method	Age	Disease	Era	Vasopressor	Sedative
Persistence	10.86	11.07	10.86	10.83	10.91
PatchTST	9.68	9.18	8.92	12.15	11.60
Chronos-2	9.55	9.12	8.86	11.70	11.35
Causal Transformer	9.72	9.30	9.05	10.45	10.28
JEPA (backbone)	9.34	10.68	9.46	11.15	11.60
Clamp	8.55	8.40	8.20	9.65	9.38
PhysioWorld	8.41	8.26	8.15	9.42	9.21

Under a shift in who the patient is, most models degrade moderately. Under a shift in how much drug is given, pure sequence models fall below persistence, while PhysioWorld stays best on every axis at the in-distribution error of its backbone.

Across databases

Bar chart of out-of-distribution skill against persistence on five cohorts under sedative shift; PhysioWorld is highest on every cohort. — Sedative intensity shift on five cohorts from MIMIC-IV, MIMIC-III and VitalDB. Skill is the fractional RMSE reduction over carry-forward persistence.

Under free-running rollout

Autoregressive RMSE against horizon for all methods; PhysioWorld has the lowest error at every horizon and the flattest growth. — Autoregressive RMSE over an eight-step horizon. PhysioWorld has the lowest error at every horizon and keeps nearly all of its skill (0.258 at h=1, 0.247 at h=8).

Does the model use the dose, and in the right direction?

Perturbing the doses at test time on the intervention-intensity slices, we count how often the predicted change has the sign pharmacology implies. This is measured on the decoded output, where the operator gives no guarantee, so it is not satisfied by construction.

The gain comes from the pharmacology, not the constraint

With the operator held fixed, skill falls monotonically as more of the stated directions are inverted, from 0.198 with the correct corpus to −0.058 when every direction is reversed, below the unconstrained backbone. Directions mined from the training records agree with pharmacology on only 48% of the pairs, no better than chance, and recover almost none of the gain. A wrong corpus is worse than none.

Dose-direction agreement

Share of eligible channel-steps whose change has the pharmacological sign when doses are zeroed (%)

Knowledge ablation

OOD skill of the same operator with different rule corpora

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