Patient-Zero: Scaling Synthetic Patient Agents to Real-World Distributions without Real Patient Data¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/thu-unicorn/Patient-Zero
Area: Medical NLP
Keywords: Virtual Patient Agents, Ab Initio EHR Synthesis, Dual-Track Cognitive Memory, NLI Logic Verification, Epidemiological Distribution Alignment
TL;DR¶
Patient-Zero presents an ab initio patient simulation framework that generates clinically realistic and diverse patient agents entirely from medical knowledge without touching real patient data, resolving the privacy-diversity bottleneck and overcoming the Stability-Plasticity Dilemma in dynamic medical dialogue through a Dual-Track Cognitive Memory System with real-time NLI verification.
Background & Motivation¶
Large language models (LLMs) have demonstrated immense promise across healthcare domains, where high-fidelity simulated patient agents offer scalable environments for medical education, Objective Structured Clinical Examination (OSCE) training, and diagnostic benchmarking. However, developing effective patient agents has long been throttled by severe privacy regulations, ethical constraints, and the scarcity of accessible clinical records. Existing patient record generation methods predominantly adhere to a derivative paradigm—relying on retrieval, fine-tuning, or generative modeling seeded by real Electronic Health Records (EHRs). Even when augmented with de-identification techniques or differential privacy noise, these derivative approaches carry persistent re-identification risks and inherently suffer from mode collapse, reproducing the demographic and geographic distribution biases of the underlying source hospital cohorts while omitting long-tail clinical presentations.
Furthermore, translating static patient records into dynamic, interactive conversational agents exposes a fundamental computational bottleneck known in cognitive science as the Stability-Plasticity Dilemma. A clinical patient agent must strictly retain factual ground truth and medical persona consistency across prolonged multi-turn consultations (stability), while simultaneously exhibiting the flexibility to naturally converse, elaborate on unscripted daily inquiries, and recall newly shared conversation history (plasticity). In practice, open-ended LLM agents frequently succumb to compliance bias and catastrophic forgetting under adversarial physician probing—conceding non-existent symptoms simply to agree with a doctor's leading questions, or contradicting their own established baseline comorbidities mid-dialogue. Conversely, rigidly scripted agents maintain stability at the expense of conversational naturalness, rendering them ineffective for realistic communication training.
To overcome both data dependency and interactive fragility, this work investigates whether realistic, multi-specialty clinical agents can be generated ab initio without exposing any real-world records. The authors hypothesize that individual patient manifestations represent causal concrete instantiations of abstract pathophysiological principles mapped onto heterogeneous demographic backgrounds. By grounding synthesis in authoritative medical knowledge bases, enforcing stratified epidemiological attribute permutations, and anchoring multi-turn dialogue in a dual-track cognitive memory gated by natural language inference, high-fidelity synthetic patient cohorts can be constructed from scratch with mathematical privacy guarantees. The core idea is to eliminate dependence on real patient records via an ab initio hierarchical synthesis pipeline governed by epidemiological priors, coupled with a Dual-Track Cognitive Memory System that enforces logical consistency and persona adherence through real-time NLI verification.
Method¶
Overall Architecture¶
The Patient-Zero architecture is organized into two tightly coupled subsystems: the Medically-Aligned Hierarchical Synthesis protocol for constructing static ground-truth patient records, and the Dual-Track Cognitive Memory System for governing dynamic, stateful agent-physician clinical consultations.
The static synthesis decomposes patient generation into an ab initio four-stage causal dependency chain: \(p(P|c) = p(E|S, B, O) \cdot p(S|B, O) \cdot p(B|O) \cdot p(O|c)\), anchored by an abstract disease concept \(c \in \mathcal{K}\). Stage I parses unstructured clinical guidelines into a standardized disease outline \(O\). Stage II conducts constrained stratified sampling across Social Determinants of Health (SDOH) dimensions to produce an epidemiological patient profile \(B\). Stage III evolves longitudinal symptom progression trajectories \(S\) based on profile risks and pathophysiological onset rules. Stage IV synthesizes granular, quantitative laboratory and imaging examination findings \(E\) while enforcing diagnosis blinding. Every stage is guarded by a deterministic and semantic verify-and-regenerate loop to prevent error propagation.
During dynamic clinical simulation, the verified record \(P = \{B, S, E\}\) is compiled into an immutable set of atomic factual propositions that populate the agent's Semantic Memory (\(M_{\text{sem}}\)). As consultations proceed, conversation events are tracked in the dynamic Episodic Memory (\(M_{\text{epi}}\)). At each dialogue turn, candidate responses generated under the injected behavioral persona \(\Psi\) are evaluated by a frozen NLI-Verifier, which enforces closed-loop memory updates: rejecting contradictions, emitting entailments invariantly, and monotonically absorbing neutral details.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Abstract Disease Concept c<br/>Authoritative Medical Knowledge Sources"] --> B["Disease Outline Standardization<br/>Structured Clinical Schema O"]
B --> C["Epidemiological Attribute Permutation<br/>Stratified SDOH Sampling into Profile B"]
C --> D["Symptom Progression Modeling<br/>Causal & Longitudinal Trajectories S"]
D --> E["Granular Clinical Expansion<br/>Diagnosis-Blinded Quantitative Exams E"]
E --> F["Dual-Track Cognitive Memory Assembly<br/>Static Semantic Memory + Dynamic Episodic Memory"]
G["Doctor Multi-Turn Query qt"] --> H["Cognitive State Candidate Generation rt<br/>Conditioned on Persona Traits Ψ"]
F --> H
H --> I["NLI Logic Verification & State Evolution<br/>Contradiction Rejection / Entailment Invariance / Neutral Expansion"]
I -->|Emit Verified Compliant Response| J["Realistic Multi-Turn Clinical Interaction"]
Key Designs¶
1. Disease Outline Standardization: Grounding Unstructured Clinical Guidelines into Structured Schemas
Raw clinical guidelines and medical encyclopedias exhibit heterogeneous syntax, varying abstraction levels, and informal descriptions that degrade downstream LLM generation. This stage acts as a medical standardization filter, deploying an LLM guided by strict ontologies, required fields, and clinical compatibility checks to transform an abstract disease concept \(c\) into a standardized JSON outline \(O\). The resulting schema explicitly formalizes core disease summaries, hallmark presentations across mild, moderate, and severe stages, demographic contraindications, special population considerations (pediatric, geriatric, obstetric), and red flag symptoms. This standardized outline serves as an immutable latent structural scaffold for all subsequent generation stages.
2. Epidemiological Attribute Permutation: Decoupled Stratified Sampling for Long-Tail Real-World Alignment
To faithfully reflect organic clinical population distributions without succumbing to generative mode collapse, this design explicitly decouples demographic attribute selection from free-form textual generation. Drawing upon the World Health Organization's Social Determinants of Health (SDOH) taxonomy, the attribute space \(\mathcal{A}\) is structured across four orthogonal dimensions: biological/demographic (\(A_{\text{bio}}\): age strata, biological sex, pregnancy status, ethnicity), socioeconomic (\(A_{\text{soc}}\): geographic setting, education, occupation, income tier), behavioral/lifestyle (\(A_{\text{beh}}\): smoking, alcohol consumption, dietary habits, physical activity level), and behavioral persona (\(A_{\text{per}}\): communication styles such as verbose, reserved, upset, and treatment preferences). Sub-vectors are sampled from categorical distributions weighted by real-world epidemiological prevalence and filtered through deterministic rejection rules:
The indicator function \(\mathbb{I}(\text{VALID}(v))\) enforces physiological and anatomical plausibility (e.g., rejecting pregnant males or pediatric-exclusive disorders in elderly strata). The resulting vector \(v = v_{\text{bio}} \oplus v_{\text{soc}} \oplus v_{\text{beh}}\) is then expanded into a coherent demographic profile \(B\), ensuring comprehensive representation of diverse and marginalized subpopulations.
3. Symptom Progression Modeling: Causal and Longitudinal Trajectory Synthesis
Standard synthetic records often present static, disconnected symptom checklists that lack clinical plausibility. This module models patient symptoms as temporal trajectories \(S\) parameterized by time of onset, progression duration, severity grades, and dynamic aggravating/relieving triggers. Symptom manifestation is conditioned causally on the patient's individual profile \(B\) (e.g., accelerated disease kinetics in immunocompromised hosts, exacerbated dyspnea linked to chronic smoking) under the constraints of disease schema \(O\). Strict chronological ordering rules enforce that symptom onset precedes formal clinical presentation, maintaining physiological causality across the longitudinal disease course.
4. Granular Clinical Expansion: Quantitative Examination Synthesis with Diagnosis Blinding
Traditional synthetic EHR tools typically generate vague qualitative remarks (e.g., "elevated white count") or prematurely reveal diagnostic conclusions, invalidating the simulation for clinical assessment. This stage generates precise, multi-modal diagnostic testing results \(E\)—including complete blood counts, comprehensive metabolic panels, arterial blood gases, microbiological cultures, and radiographic findings—conditioned on symptom profile \(S\) and target severity. The generation enforces quantitative fidelity (e.g., exact WBC \(11.6 \times 10^9/\text{L}\) with 78% neutrophils and serum CRP 38 mg/L) and strictly adheres to "diagnosis blinding," describing objective morphological and biochemical lesions while barring the explicit inclusion of the target pathology's diagnostic label. Deterministic arithmetic constraints ensure cross-variable consistency (such as concordance between reported BMI and calculated height/weight).
5. Dual-Track Cognitive Memory System: NLI-Gated Memory Evolution Resolving the Stability-Plasticity Dilemma
To eliminate compliance bias and memory drift during prolonged unscripted doctor consultations, the framework structures agent memory into declarative cognitive tiers. The static semantic memory \(M_{\text{sem}} = \{f_1, f_2, \dots, f_n\}\) decomposes the synthesized record into atomic clinical propositions to avoid lost-in-the-middle retrieval failures. The episodic memory \(M_{\text{epi}}\) tracks turn-by-turn conversational history. At turn \(t\), given cognitive state \(J_t = (M_{\text{sem}} \cup M_t^{\text{epi}}, \Psi)\) and doctor query \(q_t\), the candidate response \(r_t\) is evaluated by a frozen NLI-Verifier \(V(r_t, M_t, \Psi)\) against ground-truth facts and behavioral persona \(\Psi\). The memory transition follows a deterministic priority policy: Contradictions (or persona violations) trigger immediate candidate rejection and regeneration with corrective critique penalties; Entailments are emitted invariantly without state modification; and consistent Neutral extensions expand the episodic buffer via monotonic update:
This closed-loop gating ensures ironclad adherence to core medical ground truth while permitting flexible, persona-aligned conversational dialogue.
Loss & Training¶
The static synthesis pipeline and dynamic NLI-Verifier utilize frozen proprietary and open-source instruction-tuned LLMs without domain-specific parameter adaptation. For downstream utility benchmarking, a medical reasoning policy was trained on the synthesized Patient-Zero corpus. The training setup leveraged Qwen3-8B as both Actor and Critic backbones across a cluster of 8 NVIDIA A100 (80GB) GPUs using the VERL reinforcement learning framework and vLLM inference engine. Optimization was executed via Group Relative Policy Optimization (GRPO) with a global batch size of 128 (micro-batch size of 4 per GPU). The Actor and Critic learning rates were set to \(1 \times 10^{-6}\) and \(1 \times 10^{-5}\) respectively, with a KL divergence penalty coefficient \(\beta_{\text{KL}} = 0.001\), rollout sample size \(N = 5\) trajectories per prompt, and a maximum sequence context window of 2048 tokens.
Key Experimental Results¶
Main Results¶
The synthetic dataset comprises 60,000 comprehensive patient records spanning 98 distinct pathologies across six major clinical specialties: Cardiology, Gastroenterology, General Surgery, Neurology, Psychiatry, and Pulmonology.
The static record quality was evaluated against real-world clinical baselines (MIMIC-IV, SCRIPT, CMA Base), traditional synthetic engines (Synthea, LDP-GAN, Avatar), and open-source medical models (MERA series). Patient-Zero achieved near-perfect clinical consistency and completeness while establishing superior lexical richness:
| Method | Backbone / Source | PPL (↓) | Distinct-4 (↑) | Self-BLEU (↓) | Entity Diversity (↑) | Clinical Consistency (↑) | Clinical Completeness (↑) |
|---|---|---|---|---|---|---|---|
| MIMIC-IV | Real Clinical Data | 1.50 | 10.99 | 85.89 | 4.15 | 92.73 | 45.44 |
| SCRIPT X2B8 | Real Clinical Data | 2.38 | 51.06 | 74.75 | 0.78 | 96.12 | 56.56 |
| CMA Base | Real Clinical Data | 4.82 | 83.63 | 42.20 | 27.02 | 96.78 | 91.35 |
| Synthea | Rule-based Engine | 12.20 | 20.00 | 93.24 | 5.84 | 79.66 | 35.32 |
| LDP-GAN | Differential Privacy GAN | 10.86 | 30.70 | 91.18 | 1.10 | 85.89 | 46.40 |
| Avatar | FAMD + KNN | 9.03 | 48.43 | 70.42 | 3.74 | 95.22 | 42.84 |
| MERA-Mistral | Mistral-7B-v0.3 | 3.88 | 50.43 | 77.61 | 6.18 | 75.84 | 69.20 |
| MERA-Llama | Llama-3-70B | 3.51 | 50.57 | 77.09 | 8.27 | 82.15 | 97.00 |
| MERA-Qwen | Qwen-2.5-32B | 3.40 | 66.02 | 68.45 | 7.88 | 88.88 | 95.80 |
| Direct (w/o Hierarchy) | GPT-5 | 9.19 | 61.83 | 59.16 | 21.69 | 70.76 | 75.75 |
| Direct (w/o Hierarchy) | Gemini-2.5-Pro | 7.53 | 62.00 | 70.61 | 19.98 | 79.42 | 76.13 |
| Direct (w/o Hierarchy) | Claude-Sonnet-4 | 9.16 | 65.89 | 61.83 | 21.78 | 71.35 | 61.50 |
| Patient-Zero (Ours) | GPT-5 | 4.48 | 76.51 | 52.45 | 23.72 | 99.12 | 99.88 |
| Patient-Zero (Ours) | Gemini-2.5-Pro | 4.23 | 74.15 | 57.64 | 23.36 | 98.89 | 98.91 |
| Patient-Zero (Ours) | Claude-Sonnet-4 | 4.37 | 71.35 | 55.67 | 22.55 | 98.23 | 98.47 |
In dynamic interaction fidelity evaluations under adversarial clinical probing conducted by DOCTOR-R1, Patient-Zero decisively outperformed state-of-the-art medical agent baselines across factual stability, persona adherence, and safety robustness:
| Method | Logical Consistency (↑) | Factual Recall (↑) | Persona Alignment (↑) | Stylistic Stability (↑) | Hallucination Rate (↓) | Inducibility Resistance (↑) |
|---|---|---|---|---|---|---|
| EvoPatient (Du et al., 2024) | 87.49 | 9.78 | 80.50 | 32.95 | 5.01 | 94.87 |
| AI Hospital (Fan et al., 2024) | 97.95 | 16.68 | 78.80 | 29.33 | 6.67 | 86.36 |
| MediQ (Li et al., 2024) | 81.83 | 27.04 | 77.20 | 28.03 | 8.46 | 92.86 |
| Patient-Ψ (Wang et al., 2024) | 78.75 | 10.10 | 71.20 | 40.74 | 2.00 | 88.89 |
| Unstructured Narrative (Ablation) | 90.00 | 15.39 | 60.20 | 61.11 | 4.55 | 87.72 |
| Static Structured Schema (Ablation) | 93.18 | 14.41 | 67.20 | 64.60 | 2.73 | 79.90 |
| w/o Dual-Track Memory (Ablation) | 81.36 | 7.88 | 61.10 | 46.61 | 10.45 | 69.17 |
| Patient-Zero (Ours) | 100.00 | 33.27 | 83.70 | 75.50 | 1.82 | 97.22 |
Downstream diagnostic reasoning performance on gold-standard multiple-choice benchmarks demonstrated substantial capability gains when fine-tuning Qwen3-8B on synthetic Patient-Zero data: - Med42-v2-8B: MedQA 62.50%, MMLU 59.50% - UltraMedical-8B: MedQA 74.00%, MMLU 68.50% - Baichuan-M2-32B: MedQA 82.50%, MMLU 84.00% - Qwen3-8B (Zero-shot Base): MedQA 61.00%, MMLU 69.50% - Qwen3-8B + Real Data Training: MedQA 83.50%, MMLU 85.50% - Qwen3-8B + Patient-Zero (Ours): MedQA 85.00% (+24.0% over base), MMLU 83.00% (+14.5% over base)
In double-blind human evaluation by senior licensed physicians from top-tier hospitals (inter-rater reliability Krippendorff's \(\alpha = 0.93\)), Patient-Zero records were judged as human-authored in 60.7% of assessments (compared to 75.0% for real records). On a 5-point Likert scale, Patient-Zero achieved superior scores over real clinical notes in Information Richness (3.85 ± 0.75 vs. 3.45 ± 1.15, \(p = 0.006\)) and Completeness (4.00 ± 0.65 vs. 3.15 ± 1.27, \(p = 0.018\)), as real records frequently contained chaotic formatting and missing diagnostic sections.
Ablation Study¶
A stage-wise ablation of the static synthesis pipeline was conducted using GPT-4.1 across all six specialties (200 records generated per setting, averaging across specialties):
| Experimental Setting | Self-BLEU (↓) | Entity Diversity (↑) | Consistency (↑) | Completeness (↑) | Note |
|---|---|---|---|---|---|
| Hierarchical only | 26.64 | 54.06 | 80.46 | 97.72 | Retains hierarchy without attribute sampling or validation |
| w/o controlled sampling | 17.86 | 58.44 | 86.26 | 90.49 | Unconstrained LLM demographic generation induces mode bias |
| w/o deterministic rules | 17.92 | 57.56 | 88.60 | 95.87 | Removes physiological bounds and arithmetic checks |
| w/o semantic verifier | 21.40 | 57.72 | 84.92 | 96.36 | Omitting cross-section checking causes largest drop in consistency |
| Final-only verification | 23.00 | 56.07 | 92.26 | 97.75 | Verifying only terminal record fails to arrest cascaded errors |
| Full Patient-Zero | 15.27 | 63.33 | 95.86 | 99.55 | Integrates all hierarchical constraints and stage-wise verification |
Key Findings¶
- Dual-track memory provides structural immunity against conversational drift: Removing the Dual-Track Cognitive Memory module resulted in an 18.64% drop in logical consistency, a 25.39% decline in factual recall, and an 8.63% surge in hallucination rate, proving that standard conversational context windows fail to maintain grounding under adversarial clinical probing.
- Stage-wise verification prevents compounding error propagation: Evaluating records only at the terminal output (Final-only) underperformed stage-wise verification by 3.60% in consistency and suffered degraded Self-BLEU (23.00 vs. 15.27), demonstrating that early pathophysiological errors cascade exponentially into downstream examination findings.
- Ab initio sampling eliminates mode collapse: Goodness-of-fit hypothesis testing across all 9 demographic, socioeconomic, and lifestyle attributes confirmed perfect statistical alignment (\(p > 0.05\) across all features, global Fisher's combined test \(p = 0.86\)). In a 5,000-patient multivariate co-risk analysis, Patient-Zero accurately captured established joint dependencies, such as the amplified relative risk of hypertension in elderly obese cohorts (\(\text{RR} = 6.37\)) and fatty liver in obese drinkers (\(\text{RR} = 9.10\)).
Highlights & Insights¶
- Zero-PHI Ab Initio Synthesis: By deriving diverse patient records directly from abstract medical knowledge bases via constrained attribute permutations, the framework completely bypasses the privacy-utility trade-off, eliminating re-identification risks while delivering unprecedented diversity.
- NLI-Verifier Memory Evolution Policy: Formulating conversation memory evolution as an explicit three-way natural language inference classification (rejecting contradictions, preserving entailments, expanding neutral facts) provides an elegant, domain-agnostic mechanism to resolve the Stability-Plasticity Dilemma in goal-oriented dialogue agents.
- Standardized Diagnostic Blindness: Enforcing strict quantitative metric generation while explicitly blinding the diagnostic label provides an effective design pattern for generating clinical benchmark scenarios and examination items without leaking solutions to student models.
Limitations & Future Work¶
- Idealization Bias: The authors acknowledge that synthetic records reflect an idealized standard of clinical practice, characterized by complete fields and flawless formatting. While highly rated by physicians, this pristine structure does not mirror the typos, non-standard abbreviations, and missing documentation typical of noisy organic EHRs, leaving the impact on model noise-robustness open for investigation.
- Modality Scope & Clinical Trials: The current framework is restricted to textual records, lacking integrated radiographic imaging scans or physiological waveform signals. Furthermore, prospective randomized controlled trials assessing real-world diagnostic performance gains among resident physicians trained on Patient-Zero have not yet been completed.
- Future Directions: The team plans to incorporate native multimodal generation (synthesizing paired CT/X-ray imagery and ECG waveforms) and empirically determine optimal synthetic-to-real data mixing ratios for clinical foundation model pre-training.
Related Work & Insights¶
- vs. Traditional Synthetic Engines (Synthea, LDP-GAN): Synthea relies on rigid rule-based state transitions with restricted textual variety, while GANs trained under differential privacy collapse into blurred, generic distributions. Patient-Zero harnesses foundational LLM expressiveness to achieve far higher linguistic richness (Distinct-4 76.51 vs. 20.00/30.70) without sacrificing validity.
- vs. Derivative LLM Pipelines (MERA series): MERA prompts open LLMs using real patient record seeds, remaining tethered to source dataset distribution bottlenecks and privacy leakages. Patient-Zero constructs cases ab initio, yielding higher entity diversity (23.72 vs. 7.88) and robust alignment with nationwide epidemiological marginals.
- vs. Conversational Patient Agents (Patient-Ψ, AI Hospital): Existing patient simulators either prioritize emotional role-play at the expense of clinical factuality or rely on static scripts that suffer catastrophic compliance hallucinations under leading queries. Patient-Zero guarantees 100% logical consistency and 97.22% inducibility resistance through its real-time NLI verification gatekeeper.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ (Pioneering ab initio causal patient synthesis from abstract knowledge, seamlessly coupled with dual-track declarative memory to conquer the Stability-Plasticity Dilemma.)
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ (Comprehensive validation across 60,000 cases, 98 diseases, epidemiological statistical tests, senior physician double-blind discrimination, RL downstream training, and medical-student user studies.)
- Writing Quality: ⭐⭐⭐⭐⭐ (Rigorous formalization, beautifully structured causal chains, thorough ablation matrices, and exhaustive qualitative case studies in the appendix.)
- Value: ⭐⭐⭐⭐⭐ (Provides a privacy-safe, scalable foundation for clinical AI evaluation, synthetic EHR pre-training, and interactive medical OSCE simulation.)