Multi-Stage Patient Role-Playing (MSPRP)

• Multi-Stage Patient Role-Playing (MSPRP) is a simulation framework that generates patient responses through a staged, prompt-based process to ensure factual accuracy, medical fidelity, and persona consistency.
• The methodology decomposes the generation process into three regulated stages that sequentially enforce factual correctness, stylistic tone, and linguistic complexity.
• MSPRP enhances virtual patient simulations by addressing limitations of previous models, improving personalization, and providing actionable insights for clinical training and LLM evaluation.

Multi-Stage Patient Role-Playing (MSPRP) is a technically rigorous framework for simulating highly realistic and personalized clinical interactions by decomposing patient response generation into regulated, sequential stages. MSPRP, as formalized in recent research, offers a prompt-driven, training-free approach that directly addresses the limitations of prior LLM-based and template-driven virtual patient simulations with respect to medical fidelity, persona consistency, and contextual naturalness (Jiang et al., 16 Jan 2026). The architecture, variants, and impact of MSPRP span both educational and LLM evaluation domains and have been instantiated across multiple frameworks for language, agent-based, and adaptive simulation systems (Lee et al., 31 May 2025, Du et al., 2024, Marez et al., 20 Dec 2025).

1. Formal Problem Definition and Motivations

The primary objective of MSPRP is, given a structured patient persona $P$, clinical content/context $C$, dialogue history $H$, and a new doctor prompt $Q$, to generate a patient utterance $A$ that is (i) factually and temporally coherent with $C$ and $H$, (ii) persona-consistent with an individualized, multi-dimensional $P$, and (iii) contextually and linguistically appropriate. Key deficiencies in existing approaches include excessive uniformity and lack of personality in LLM outputs, reliance on homogeneous or LLM-generated template data, and inconsistent persona representation due to unstructured attribute modeling (Jiang et al., 16 Jan 2026). MSPRP explicitly targets these deficits by embedding a structured, multi-dimensional patient persona and decomposing the generation process into three regulated stages.

2. MSPRP Pipeline: Stagewise Decomposition

The MSPRP pipeline operates via three sequential, fully prompt-based subroutines, each regulating distinct constraints on the output. At each stage, specific elements of the patient persona $P$ are activated to enforce medical, stylistic, and expressive requirements.

Stage	Input Parameters	Function	Output Constraint
Basic Generation	$P$, $C$, $H$, $Q$	Factually correct, context-anchored answer (no stylistics)	Completeness, consistency, accuracy
Style Injection	$P$ (personality, emotion), scenario ID, $A_1$	Injects stable communication style, tone, and emotion	Persona-dependent style
Expression Consistency	$P$ (history recall, comprehension, fluency), $A_2$, $H$	Regulates linguistic complexity, recall behavior, term usage	Expressive capacity, naturalness

In practice, this workflow can be summarized by the functional composition $A = G(P, C, H, Q)$ with $G$ implemented as the above pipeline (Jiang et al., 16 Jan 2026).

3. Patient Persona Structure: Five-Dimensional Attribute Model

Each patient persona $P$ is a 5-tuple:

$$P = [p^\text{Personality},\ p^\text{Emotion},\ p^\text{HistoryRecall},\ p^\text{Comprehension},\ p^\text{Fluency}]$$

• $p^\text{Personality}$ (e.g., anxious, cooperative)
• $p^\text{Emotion}$ (e.g., fearful, calm)
• $p^\text{HistoryRecall}$ (low, medium, high)
• $p^\text{Comprehension}$ (low, medium, high)
• $p^\text{Fluency}$ (poor, average, fluent)

Communication style is modulated by $p^\text{Personality}$ and $p^\text{Emotion}$, while expressive capacity is regulated by the remaining dimensions. Each stage in the MSPRP pipeline dynamically applies relevant persona parameters: Stage 1 calibrates recall, Stage 2 determines stylistic features, and Stage 3 enforces linguistic complexity and terminological fidelity (Jiang et al., 16 Jan 2026).

4. Instantiations and Extensions Across Frameworks

Multiple agentic and adaptive systems ground their patient simulation protocols in MSPRP or close analogs.

Adaptive-VP implements stages as behavioral modes indexed by dynamic trainee assessment; transitions are governed by real-time scores assigned to trainee utterances. The framework uses modular LLM orchestrations and scenario-driven symbolic governance for both response calibration and safety (Lee et al., 31 May 2025).

EvoPatient operationalizes MSPRP as a coevolutionary loop: discrete dialogue stages (chief complaint, triage, interrogation, and conclusion) are iteratively refined as doctor and patient agent populations exchange validated dialogue fragments. Fitness and selection are based on strictly coded criteria for both answer and question quality, with continual updating of prompt libraries for both roles (Du et al., 2024).

Agentic AI Framework for General Practitioner Training decomposes its process into vignette generation (scenario control), persona-driven dialogue (interaction control), and assessment/feedback (critique control). Persona parameters (largely Big Five trait vectors) are explicitly mapped to persona text used in dialogue prompts, and standards-based rubrics govern post-session feedback (Marez et al., 20 Dec 2025).

5. Evaluation Methodologies and Benchmarking

Performance of MSPRP is quantified via both automatic metrics and human-aligned ratings:

• Automatic metrics: BLEU-$n$, ROUGE-L, METEOR, and BERTScore, capturing $n$-gram, sequence, and embedding-level semantic alignment (Jiang et al., 16 Jan 2026).
• Human-model evaluations: State-of-the-art LLMs score outputs (1–5) for persona consistency, factual accuracy, naturalness, and contextual relevance.

Key observed improvements (Qwen2.5-7B, MSPRP vs baseline):

Metric	Baseline	+MSPRP	Relative Gain
BLEU-4	0.0329	0.0375	+13.8%
ROUGE-L	0.2051	0.2196	+7.1%
Persona Consistency	3.748	3.905	+0.157 (5pt)
Factual Consistency	3.795	3.918	+0.123 (5pt)

In adaptive and agentic frameworks, additional validation includes inter-rater agreement ($\kappa>0.88$ for tone, de-escalation, and prohibited behaviors (Lee et al., 31 May 2025)), as well as user-perceived consistency and realism in multi-user studies (Marez et al., 20 Dec 2025).

6. Limitations and Open Challenges

While MSPRP substantially increases persona fidelity and response diversity, key limitations remain:

• Rule engineering: The Interaction Matrix and prompt design require significant domain expertise; generalization to new specialties or languages entails manual redesign (Jiang et al., 16 Jan 2026).
• Scenario identification: Rule-based scenario assignment can misclassify turn function, impacting style rendering (Jiang et al., 16 Jan 2026).
• Generality: Empirical validation has focused primarily on Chinese language gastroenterology; cross-linguistic or multi-specialty adaptability is not yet demonstrated (Jiang et al., 16 Jan 2026).
• Extensibility: Pipeline design facilitates modular extension, yet full integration of reinforcement learning, multi-agent simulation, and internal persona-conditioning remains a direction for future work.

7. Representative Outputs and Use Cases

MSPRP enables nuanced, contextually coherent patient responses reflecting differentiated personality and cognitive traits. Example (Qwen2.5-7B, anxious-paranoid persona, high recall/comprehension):

"I’m really worried that delaying the gastroscopy won’t identify what’s causing my pain. Last time, you said the findings could explain why my stomach burns at night—if we wait, might we miss something important? I’d feel more at ease knowing exactly what’s wrong as soon as possible."

This contrasts with the neutral, bland baseline response, demonstrating explicit persona expression, contextual recall, and appropriate use of medical terminology (Jiang et al., 16 Jan 2026).

In adaptive training, MSPRP-like protocols facilitate dynamic escalation or de-escalation (e.g., from hostility to cooperation) based on real-time trainee performance scoring, which mirrors pedagogically valuable real-world clinical interaction patterns (Lee et al., 31 May 2025).

• "Multi-Stage Patient Role-Playing Framework for Realistic Clinical Interactions" (Jiang et al., 16 Jan 2026)
• "Adaptive-VP: A Framework for LLM-Based Virtual Patients that Adapts to Trainees' Dialogue to Facilitate Nurse Communication Training" (Lee et al., 31 May 2025)
• "LLMs Can Simulate Standardized Patients via Agent Coevolution" (Du et al., 2024)
• "An Agentic AI Framework for Training General Practitioner Student Skills" (Marez et al., 20 Dec 2025)