Patient Simulators That Know What They Don't Know

AI doctors are increasingly evaluated, and even trained, by talking to simulated patients. To stop these simulated patients from making up symptoms, recent systems tie their answers to a medical knowledge graph, a database of facts such as "pneumonia causes cough." This creates a hidden flaw. A graph lists what is known, not everything that is true. A graph-grounded patient therefore tends to answer "no" to any question about a finding the graph does not mention. The doctor agent that asks about a real but unlisted symptom is told it is absent, and is graded as if it asked a useless question. The flaw is worst for imaging. Medical graphs link text concepts and almost never record what appears on an X-ray. A doctor that orders a chest radiograph is asking about exactly the region the graph cannot see.

We propose to measure this closed-world bias and to remove it. Our fix is an open-world patient simulator, built around three ideas. Each fact has three possible states: present, absent, or unknown. Present and absent are decided by the patient's own real multimodal record, meaning the radiograph, the report, labs and notes. The knowledge graph is consulted only when the record is silent, and only to supply plausible detail, never to issue a denial. Questions whose answer is unknown are neither rewarded nor penalized. The central claim is testable. Closed-world simulators distort the ranking of doctor agents. When used as a reinforcement-learning sandbox (a trial-and-error training loop where the agent is rewarded by the simulator's responses), they train policies that stop asking about off-graph findings and under-order imaging. An open-world simulator fixes both problems.

What exists and what is missing

The closest work is MedKGEval. It adds a "Director" agent that retrieves subgraphs from a merged medical graph (CMeKG, PrimeKG and others) to steer a patient agent. It reports that patient factual consistency rises from roughly 60% to over 91%. That result motivates us, because grounding clearly reduces invented symptoms. It also leaves the problem we target. The file notes that MedKGEval assumes the graph covers all required facts, is text-only, and is used only for evaluation.

Doctor-R1 trains an 8B doctor with a group-relative policy-gradient method against an LLM patient built from 100,000 KaMed-derived dialogues. That patient is anchored to cases, not to a graph. The paper also flags evaluator hacking as a risk. Salus frames diagnosis as sequential test ordering, but its test results are text snippets produced by an LLM. It cannot hand the agent a real image.

Case-anchored simulation by itself is therefore not new. What none of these systems do is separate documented-absent from undocumented. Patient records are incomplete too: a case simulator that answers "no" whenever the record is silent has the same closed-world bias as a graph-grounded one.

Two strands of graph work constrain the design.

  • Strict decoding blocks missing facts. Graph-constrained Reasoning (GCR) forces generation along paths that exist in the graph. It is fully faithful, but it cannot output a fact that is missing from the graph.
  • Loose decoding hallucinates. DialogGSR falls back to the model's own memory when subgraphs have gaps, and hallucination returns.

KMQA quantifies the stakes in medical QA: 44% of its errors came from information missing in both text and graph. Our three-valued design is a middle path between these failure modes. Facts are pinned to evidence when evidence exists. Silence is expressed as patient uncertainty ("I'm not sure, I haven't noticed") rather than as a denial or an invention.

The multimodal side reuses existing tools rather than inventing them.

  • Graph incompleteness on the image side. PrimeKG has 129,375 nodes and over 61,000 contraindication edges, but no links to image features. That is precisely the incompleteness we expect to bias imaging decisions.
  • Three-valued labels from reports. MIMIC-CXR comes with NegBio/CheXpert-extracted labels that already distinguish positive, negative and uncertain findings. RadGraph turns reports into anatomy–observation graphs, with 0.94 entity F1 and 0.82 relation F1. Both give per-patient structure that we can use as a three-valued record.
  • Image interpretation for the doctor. RadAgents supplies tool-based X-ray interpretation for the doctor side.

Research Program

We first build the case substrate. We take MIMIC-CXR studies with their reports and labels and extract RadGraph structures. Where the studies can be linked to the same patients' notes and lab series in MIMIC-IV (the EHR setting used by ClinRaGen), we attach history and labs. Each case becomes a record of facts tagged present, absent or silent. Facts carry their modality: some are stated by the patient, some are measured, and some are visible on the image.

How the simulator answers. Free-text questions go to a fact matcher. If the record has the fact as present or absent, the patient answers accordingly, and the natural-language surface is left unconstrained so that vagueness and colloquial speech remain. This addresses the tension between strict fact constraints and natural dialogue. If the record is silent, the simulator checks whether PrimeKG links the case's diagnosis to that finding.

  • Graph link exists: the patient gives a hedged, low-commitment answer.
  • No graph link: the patient expresses ignorance.

Ordered tests. If the doctor orders a chest X-ray, it receives the actual image. A text report is offered only as a separate ablation condition.

How the reward is scored. Turns about present or absent facts are scored. Turns about unknown facts carry a small fixed question cost and nothing else. Three things prevent the agent from exploiting the "unknown" state:

  • the fixed question cost on unknown turns;
  • a cap on the number of turns;
  • an outcome reward computed against the held-out record, never against the simulator's own outputs.

Measuring how incomplete the graph is. Before any agent is trained, we measure, over the substrate, the fraction of record-documented findings for each diagnosis that PrimeKG does not link to it. We report this separately for text-type findings and image-type findings. Our hypothesis is that the image-side gap is much larger. This analysis is cheap, and it decides whether the rest of the program matters.

Baselines

Closed-world graph-grounded simulator (naive reference). We reimplement MedKGEval's Director-grounded patient over PrimeKG on our cases. Anything not retrieved from the graph is answered as absent. This is the strongest existing competitor because its reported jump in factual consistency, from about 60% to over 91% (MedKGEval), is the best grounding result in the file. The doctor agents, case pool and turn budget are identical across simulators. Beating it means two things:

  • statistically fewer false denials, i.e. "no" answers about findings present in the record;
  • no loss in factual consistency on record-documented facts, measured by the same consistency protocol.

Case-anchored simulator without a graph. This is a Doctor-R1-style LLM patient prompted with the record, where silence defaults to "no." It tests whether case anchoring alone suffices. No reported result found on Lacuna for patient-side false-denial rate.

Graph-plus-memory fallback. This is DialogGSR-style behavior, where the patient falls back to the LLM's own knowledge when the graph is silent. We measure its rate of contradicting the record.

Doctor-side references. The policy is trained in the same sandbox under each simulator. We use Salus-7B and Doctor-R1-8B scale backbones with an identical RL compute and tuning budget across conditions. We report published reference points: Salus 83.64% Top-1 on CompDiag-Bench high-complexity cases (Salus), Doctor-R1 36.29 on HealthBench (Doctor-R1), and RadAgents 73.6% on ChestAgentBench (RadAgents). These anchor that our doctor agents are competent. They are not direct competitors, since none was measured on our substrate.

Ablations. These isolate the key ingredient by removing one piece at a time:

  • the three-valued unknown state, forcing binary answers;
  • the graph's role in the silent case;
  • the raw image, replacing it with report text.

Why not just…?

Why not just tell the LLM patient to say "I don't know" when unsure? The model has no reliable signal for when it should be unsure. It cannot tell an undocumented finding from a documented absence without the record's explicit negative labels, and freed from constraints it reverts to its own memory, as DialogGSR's failure mode shows. We include this prompt-only variant as a baseline and measure both its false denials and its contradictions with the record. If it matches our simulator, the structured three-valued record is unnecessary, and we would report that.

Why not just complete the knowledge graph first? This is a real threat. We test it by adding to PrimeKG the disease–finding edges mined from RadGraph over the training cases, and rerunning the closed-world simulator. Two things limit it. Completion yields disease-level edges, not patient-level facts. And it cannot create documented absences. We expect it to narrow the false-denial gap but not close it. The residual gap on held-out cases is the deciding number.

Why not evaluate on retrospective records with no simulator at all? Static records cannot answer a question the original clinician never asked. Wang et al. argue that static evaluation of interactive systems misstates their quality. MedKGEval found failures such as premature convergence only in turn-by-turn interaction. RL training also requires an interactive environment.

Decisive Experiments

1. Ranking fidelity. We evaluate a fixed pool of doctor agents under each simulator. These include frontier models prompted directly, a Salus-style controller and Doctor-R1-style policies. We compute the rank correlation of each simulator's scores with an oracle score from the held-out record: final diagnosis correctness plus the fraction of documented positive findings elicited. The open-world simulator should correlate more strongly. The gap should be largest for agents that order imaging.

2. Dose response. We randomly delete increasing fractions of PrimeKG disease–phenotype edges. Closed-world rewards should fall roughly in proportion to the edges removed, since each deleted edge converts a true answer into a denial. Open-world rewards should stay nearly flat.

3. Training consequence. This is the costliest and most important experiment. We train identical policies under the closed-world and open-world simulators, then test both on held-out record-anchored cases with real images. The prediction is that the closed-world-trained policy learns to ask fewer off-graph questions, orders fewer radiographs, and loses diagnostic accuracy. If the two policies end up indistinguishable, the closed-world bias is a curiosity for evaluation and not a training hazard, and the proposal's main claim fails.