Decoupling Patient Personas and Clinical Workflows: A Dual-Agent Approach to Realistic Medical Dialogue Generation

• Pre-consultation dialogues gather comprehensive patient information for triage and diagnostic decisions, while privacy, ethical constraints, and expert-annotation costs make large-scale collection difficult.Source: §1 · p. 1
• Existing approaches trade scalability against reliability: fine-grained control can require expert-designed rules and manual verification, while iterative multi-turn control can increase complexity and reduce consistency.Prompt-driven methods may have limited clinical reliability; workflow-driven generation can require costly expert verification, and intent-driven simulation can require repeated LLM calls.Source: §1 · p. 1, Figure 1 · p. 1
• CoRPS-Dial decouples patient heterogeneity from guideline-aligned process modeling through interactions between a persona-driven patient agent and a process-guided doctor agent.Personas are constructed from de-identified cardiology consultations, while process rules encode workflow structure and clinical priorities.Source: §1 · p. 2
• Its structured generation and post-hoc quality-control pipeline is designed to ensure medical consistency, workflow completeness, and realistic patient heterogeneity.Source: §1 · p. 2
• A CHD case study finds CoRPS-Dial dialogues better aligned with real-world pre-consultation dialogues in structure, workflow adherence, and coherence; downstream fine-tuning demonstrates their utility for medical dialogue tasks.Source: §1 · p. 2

2 Related Work

• Medical dialogues include real doctor–patient interactions as well as dialogues produced through human annotation or model-based synthesis.Source: §2 · p. 2

2.1 Real-World Medical Dialogues

• Real-world dialogue resources span medical domains and scales, but many contain general or weakly structured interactions rather than disease-specific pre-consultation workflows.Privacy and ethical constraints also limit the availability, scalability, and reusability of authentic dialogues.Source: §2.1 · p. 2

2.2 Synthetic Medical Dialogues

• LLM-generated dialogues offer greater controllability, lower privacy risk, and scalability, but remain vulnerable to hallucinations and weak clinical grounding.Prior work ranges from medical Q&A and simulated consultations to multi-turn interactions.Source: §2.2 · p. 2
• Rule-based and hybrid quality-control approaches seek clinical plausibility, yet may depend on expert involvement, lack guideline alignment, or constrain consistency during generation.CoRPS-Dial responds by combining patient personas with guideline-aligned workflow rules.Source: §2.2 · p. 3

3 Methodology

• The methodology presents CoRPS-Dial as a dual-agent framework for controllable generation of clinically grounded multi-turn dialogues.The framework separates workflow decisions from their language realization.Source: §3 · p. 3, §3.4 · p. 4

3.1 Problem Formulation

• The framework decouples disease-specific workflow rules from disease-agnostic patient profiles to control procedural completeness, clinical consistency, and patient variability.A patient instance governs responses, while workflow rules encode inquiry structure and priorities.Source: §3.1 · pp. 3–4

3.2 Patient Persona Construction

• Patient profiles represent demographics, symptoms, emotions, communication traits, and narrative source, and controlled perturbations expand the profile space.Profiles are extracted from de-identified dialogues; age varies within ±5 years, while other attributes are sampled from their attribute sets.Source: §3.2 · p. 4

3.3 Clinical Process Modeling

• A guideline-grounded inquiry tree preserves mandatory items and safety constraints while adding optional branches mined from real consultations.The doctor agent prioritizes mandatory, high-risk, and prerequisite nodes, then realizes selected inquiries as natural-language questions.Source: §3.3 · p. 4, §3.4 · p. 4

3.4 Dual-Agent Dialogue Generation

• The doctor agent selects inquiries from the workflow tree, while the patient agent responds using the inquiry, dialogue history, and fixed profile.Dialogue ends when mandatory nodes are covered or guideline-defined stopping criteria are met.Source: §3.4 · pp. 4–5

3.5.1 Hard & Soft Quality Control

• Hard-first quality control flags role, process, contextual-coherence, and text-quality violations before GLM-4-9B scores dialogues for refinement.Dialogues below the average reward of real dialogues under the same dimensions are selected for refinement.Source: §3.5.1 · p. 5

3.5.2 Selective Utterance Revision

• Selective revision targets flagged utterances and replaces an original only when the revised utterance earns a higher reward score.Revision uses the utterance, dialogue context, patient profile, and assessment feedback rather than regenerating the full dialogue.Source: §3.5.2 · p. 5

4.1 Datasets

• The dataset comprises 256 real pre-consultation dialogues and 9,714 synthetic dialogues produced from 5,120 augmented profiles after quality control.The real dialogues cover complaints, present-illness history, and risk factors; downstream data are split 8:1:1.Source: §4.1 · p. 5

4.2 Evaluation Metrics

• Evaluation covers dialogue structure, semantic consistency, lexical diversity, human and LLM ratings, and downstream dialogue performance.Table 1 compares structure, semantics, and diversity; Table 2 reports role consistency, process consistency, contextual coherence, and text quality.Source: §4.2 · p. 5, Table 1 · p. 6, Table 2 · p. 6

4.3 Implementation Details

• Implementation uses Qwen3-32B for both agents and persona and workflow construction, with post-hoc quality control and selective rewriting.The rewrite threshold is θ = 7.2; the experiments also fine-tune three models with LoRA and use specified LLM and embedding models for evaluation.Source: §4.3 · p. 6

4.4 Baseline

• Generation baselines remove workflow modeling, persona conditioning, dual-agent separation, or quality control and selective rewriting; downstream comparisons include each model before and after fine-tuning.Table 3 is captioned as comparing synthetic and real dialogue quality across structure, semantics, and diversity.Source: §4.4 · pp. 6–7, Table 3 · p. 7

5.1 Main Results

• CoRPS-Dial consistently outperforms all baselines across structural, semantic, and diversity metrics, and performs best in human and DeepSeek-v2 evaluation.The reported comparisons attribute shorter or less coherent dialogues to Raw Doctor, reduced richness to Free Patient, less diversity to Single Agent, unstable roles to Minimal Agents, and more verbosity and textual inconsistency to No QC/SR.Source: §5.1 · p. 7

5.3 Human and LLM Evaluation of Real and Synthetic Dialogue Quality

• Figure 3 compares human-expert and LLM-as-a-Judge ratings of 50 real and 50 synthetic dialogues on a 10-point Likert scale.Human ratings average three domain experts; LLM ratings use Deepseek-v2 across three epochs.Source: Figure 3 · p. 7

5.2 Quality and Scalability of Synthetic Dialogues

• Synthetic dialogues retain stable structural and semantic quality across corpus sizes, although larger scale brings mild surface-level repetition.The authors report that lexical richness and semantic coherence are preserved across the sampled sizes.Source: §5.2 · p. 7
• Across 5,120 stylistically augmented dialogues, statistical and semantic consistency and lexical diversity are preserved without systematic chief-complaint repetition or distortion.The set includes one representative dialogue per portrait of the augmented patient profile.Source: §5.2 · pp. 7–8

5.4 Downstream Validation

• Supervised fine-tuning with synthetic data improves downstream task performance across most evaluated models, with gains in coverage, efficiency, and dialogue control.The authors note significant enhancements for LLaMA-8B and DeepSeek-R1-14B, and say benefits are maximized with model-specific fine-tuning strategies.Source: §5.4 · p. 8

5.5 Case Study

• Compared with existing methods, CoRPS-Dial produces more consistent roles, coherent interaction flows, and cleaner language in qualitative dialogue comparisons.The cited issues in other methods include insufficient symptom inquiry, limited persona diversity, unstable interaction patterns, and textual inconsistencies.Source: §5.5 · p. 8

6 Conclusion

• CoRPS-Dial combines structured workflows, personalized patient personas, dual-agent interaction, and two-stage post-hoc quality control.The conclusion says the method addresses role inconsistency, process redundancy, semantic drift, and textual noise in synthetic dialogue generation.Source: §6 · p. 8
• Evaluations show CoRPS-Dial outperforms all baselines, producing coherent, clinically faithful, and diverse dialogues.Case studies further demonstrate improved role consistency and workflow fidelity.Source: §6 · p. 8
• The authors describe CoRPS-Dial as a scalable, reliable approach to constructing clinical dialogues for future controllable, data-centric medical conversational research.Source: §6 · p. 8

Limitations

• LLM hallucination remains difficult to eliminate; the authors suggest expert validation or external knowledge as possible reliability improvements.Source: Limitations · p. 9
• Validation currently focuses on CHD, so generalizability to other diseases remains to be assessed.Source: Limitations · p. 9
• Single-institution data may introduce bias; the authors suggest multi-center data to enhance diversity.Source: Limitations · p. 9

Ethical Considerations

• Participants provided informed consent, and transcripts were de-identified before processing to protect privacy.Source: Ethical Considerations · p. 9
• Audio transcription used a locally deployed system, while transcripts and metadata were stored on encrypted institutional servers with restricted access.Source: Ethical Considerations · p. 9
• Synthetic dialogues are intended for research and methodological evaluation, not direct clinical decision-making without professional oversight.The authors cite potential generative-model inaccuracies or hallucinations as the reason for this boundary.Source: Ethical Considerations · p. 9

A.1 Noise Handling and Structural Normalization of Real-World Clinical Dialogues

• Filtering removes filler words and clinically irrelevant self-corrections but preserves repeated symptom mentions when they convey emphasis.Source: §A.1 · p. 11
• Turns consisting only of greetings, social interaction, or unrelated chit-chat are discarded; utterances under three tokens without medical entities are considered for removal.Source: §A.1 · p. 11
• Clinically relevant tokens, symptom mentions, and temporal indicators are retained, and filtering does not modify the patient's reported meaning.Source: §A.1 · p. 11

A.3 Details of Evaluation Metrics

• The evaluation metrics cover dialogue structure, semantic consistency, lexical diversity, and downstream information gathering, question quality, and consultation completion.Downstream measures also include independent LLM ratings of coherence, relevance, clinical logic, informativeness, and fluency.Source: §A.3 · p. 11

A.4 Downstream Evaluation Metrics

• Downstream dialogue performance is evaluated for medical-information coverage, turn efficiency, use of recent patient responses, repetition, invalid questions, and appropriate termination.An independent LLM evaluates overall consultation quality across coherence, relevance, clinical logic, informativeness, and fluency.Source: §A.4 · pp. 11, 13

A.2 Scoring Criteria for Reward Evaluation

• GLM-4-9B scores synthetic dialogue quality across role consistency, process consistency, contextual coherence, and basic text quality, averaging the criteria.Source: §A.2 · p. 11, Table 5 · p. 12

A.5 Quality and Scalability Analysis of Synthetic Dialogues

• Profile-based expansion introduces natural variability while preserving clinical fidelity, and small expanded subsets reflect full-dialogue quality without bias from repeated core information.Across scales, dialogues maintain about 42 turns and high semantic similarity; lexical entropy remains stable as dialogue volume grows.Source: §A.5 · p. 13, Table 6 · p. 14

A.6 Human–LLM Agreement Analysis

• Human–LLM Cohen’s kappa shows strong consistency across the evaluation dimensions.Source: §A.6 · p. 14, Figure 5 · p. 14

A.7 Case Study

• Baselines show distinct shortcomings: raw doctors omit targeted symptom and risk-factor inquiry, free patients give template-like responses, and minimally constrained agents produce unfocused dialogues.Without post-hoc quality control, structured generation still has role inconsistency, incomplete or redundant steps, and textual noise; refinement yields more consistent roles, coherent flows, and cleaner language.Source: §A.7 · p. 14