When to Trust Simulated Patient Responses:Support-Gated World Models for Vasopressor Decision Support
Organizations: Sun Yat-sen University, Guangzhou, Guangdong, China
Abstract
Clinical world models expose language-model agents to predicted patient responses, but simulations can be unreliable for poorly supported treatments or unfamiliar hospitals. We introduce Support-Gated World-Model decision support, a lightweight extension of a propose--simulate--refine agent. A recurrent dynamics ensemble predicts short-horizon physiological responses to documented norepinephrine dose categories, while a behavior model estimates observational action support. The interface withholds predictions with insufficient support or excessive ensemble disagreement and abstains when no candidate qualifies. Evaluation on eICU-CRD includes 70 hospitals and 30,700 decision transitions, with 14 hospitals held out for testing. At 62.0% factual-transition coverage, joint gating yields a MAP mean absolute error of 7.3 mmHg, compared with 9.2 mmHg for random selection at the same coverage. The gated agent achieves 60.0% agreement with recorded actions at 81.0% decision coverage, compared with 55.0% agreement and 99.5% coverage for an ungated ensemble agent. Component ablations show lower factual risk with joint gating than with either gate alone.These results characterize a reliability--coverage trade-off in selective clinical simulation; they do not identify individualized treatment effects or establish improved patient outcomes.
Figures & tables
| Characteristic | Training | Validation | Test |
| Hospitals | 42 | 14 | 14 |
| Unique patients | 5,200 | 1,650 | 1,800 |
| Eligible ICU stays | 5,400 | 1,730 | 1,880 |
| Decision transitions | 18,500 | 5,900 | 6,300 |
| Documented zero-dose transitions | 210 | 65 | 70 |
| Missing MAP targets (%) | 9.4 | 10.1 | 11.0 |
| Selection rule | Coverage (%) | MAP MAE | HR MAE |
|---|---|---|---|
| All factual transitions | 100.0 | 9.2 | 11.6 |
| Support only | 77.0 | 8.6 | 10.9 |
| Disagreement only | 79.0 | 7.9 | 10.3 |
| Joint gate | 62.0 | 7.3 | 9.6 |
| Random selection (matched) | 62.0 | 9.2 | 11.6 |
| Method | Coverage (%) | Agreement (%) | Unsupported (%) | Calls/case |
|---|---|---|---|---|
| Behavior cloning | 100.0 | 53.0 | 0.0 | 0.0 |
| Direct LLM | 98.0 | 48.0 | 14.0 | 1.0 |
| Ungated ensemble agent | 99.5 | 55.0 | 12.0 | 4.8 |
| SG-WM | 81.0 | 60.0 | 0.0 | 4.6 |
| Interface | Support gate | Disagreement gate | Factual risk | Agent coverage (%) |
|---|---|---|---|---|
| Ungated | No | No | 0.53 | 99.5 |
| Support only | Yes | No | 0.50 | 93.0 |
| Disagreement only | No | Yes | 0.46 | 90.0 |
| Joint gate | Yes | Yes | 0.43 | 81.0 |
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| Component | Configuration |
|---|---|
| Dynamics ensemble | 3 independently initialized GRUs; 2 layers per GRU; 128 hidden units; dropout 0.1 |
| Dynamics optimizer | AdamW; learning rate ; batch size 128; at most 100 epochs; early-stopping patience 10 |
| Stability-assessment seeds | 13, 42, 2026 |
| Behavior support estimator | Separate GRU classifier with categorical cross-entropy loss |
| Action representation | Documented zero dose plus 4 positive-dose bins; training cutpoints , , g/kg/min |
| Proposal and refinement model | Qwen2.5-7B-Instruct |
Clinical Decision Support
Clinical Decision Support uses data, medical knowledge, and AI to help clinicians assess patients and choose tests, diagnoses, and treatments.
See all 256 papers