Generating High Quality Synthetic Data for Dutch Medical Conversations

Cecilia Kuan; Aditya Kamlesh Parikh; Henk van den Heuvel

cs.CLMar 25, 2026

Generating High Quality Synthetic Data for Dutch Medical Conversations

Authors: Cecilia Kuan, Aditya Kamlesh Parikh, Henk van den Heuvel

Organizations: Center for Language Studies, Radboud University, Nijmegen, Netherlands

Abstract

Medical conversations offer insights into clinical communication often absent from Electronic Health Records. However, developing reliable clinical Natural Language Processing (NLP) models is hampered by the scarcity of domain-specific datasets, as clinical data are typically inaccessible due to privacy and ethical constraints. To address these challenges, we present a pipeline for generating synthetic Dutch medical dialogues using a Dutch fine-tuned Large Language Model, with real medical conversations serving as linguistic and structural reference. The generated dialogues were evaluated through quantitative metrics and qualitative review by native speakers and medical practitioners. Quantitative analysis revealed strong lexical variety and overly regular turn-taking, suggesting scripted rather than natural conversation flow. Qualitative review produced slightly below-average scores, with raters noting issues in domain specificity and natural expression. The limited correlation between quantitative and qualitative results highlights that numerical metrics alone cannot fully capture linguistic quality. Our findings demonstrate that generating synthetic Dutch medical dialogues is feasible but requires domain knowledge and carefully structured prompting to balance naturalness and structure in conversation. This work provides a foundation for expanding Dutch clinical NLP resources through ethically generated synthetic data.

Figures & tables

Figure 1: Broader Project Workflow - this study focuses on the Text Dataset component (highlighted)

Figure 2: Synthetic Dialogue Text Generation Workflow

Figure 3: Synthetic Dialogue Statistics - Number of Words and Turns. D1–D9 denote individual-generated dialogues.

Metric	Mean	SD
Alternation rate	0.973	0.021
Role consistency	0.012	0.007
ASL	16.18	4.03
Average SPT	2.14	0.37
Topic coverage	0.889	0.132
TTR	0.377	0.04

Table 1: Quantitative Evaluation Summary

Figure 4: Role Consistency - Average per Dialogue

Figure 5: Role Consistency - Roles

Figure 6: Scores: ASL and SPT

Figure 7: Topic Coverage - Proportion per Dialogue

Figure 8: Scores: TTR & MSTTR

Figure 9: Human Evaluation - Scores per Category

Figure 10: Human Evaluation - per Rater in each Category

Figure 11: Human Evaluation - Inter Rater Reliability

Measure	D9	D3	D6
Length	610	972	1266
MSTTR	0.85	0.825	0.843
TTR	0.44	0.384	0.321
ASL	10.97	11.61	14.79
Turns	26	61	50
SPT	2.44	1.62	2.44

Table 2: Comparison of Selected Dialogues by Key Metrics

Figure 12: Spearman Correlation ( $\rho$ ) - Qualitative vs Quantitative Scores

Coherency
Points	Definition and Examples
1-2	Frequent order/timing mistakes (non-sequential turns, confusing switches, inconsistent tense). Dialogue feels unnatural and hard to follow. Ex - doctor and patient turns are frequently mixed up, inserted after goodbye:: "Arts: Goed, laten we nu eens kijken naar wat u al doet en wat we kunnen verbeteren."
3	Mostly sequential and moderately coherent, but some abrupt or unnatural/unclear transitions between topics (e.g. sudden topic change, unclear turn boundaries). Ex - "Patiënt: Ik probeer elke twee uur een half kopje te drinken. Arts: En wat betreft uw voeding?"
4-5	Logically flows, correct turn-taking, consistent tenses, naturally progressing conversation; mostly coherent with minor jumps between topics without clear transition.. Ex - "Patiënt: Ik probeer elke twee uur een half kopje te drinken. Arts: En wat betreft uw voeding?"

Table 15

Consistency
Points	Definition and Example
1-2	Frequent factual errors, hallucinated medical info, missing or contradictory details with previous statement and context. Lacks accuracy in medical facts, treatment options, or role actions. Ex - contains hallucinated diseases/treatment - "Arts: Je zegt dat je allergisch bent en Diazepam vijf keer per week neemt."
3	Some inconsistencies, minor errors or missing pieces, e.g., partial topic coverage, minor fact mistakes, not linking blood pressure to kidney function after discussing both. Ex - "Patiënt: Ik probeer mijn waterinname te verhogen, maar soms vergeet ik." (General lifestyle is relevant, but lacks specific detail or contradicts earlier plan.)
4-5	Consistent and accurate; mostly or all facts align with scenario and medical knowledge, topics covered.

Table 16

Fluency
Points	Definition and Example
1-2	Use of Dutch is awkward or ungrammatical, regardless of content. Poor quality sentences, long sentences, repetition in sentences. Translations-like errors. Ex - "Het rug pijn is niet goed voelen en ik zijn moe altijd is van pijn."
3	Readable with occasional awkwardness, odd phrasing, or grammar slips; generally understandable. Ex - "Ja, ik ben nemen de pillen en ik voel minder, maar niet blij altijd."
4-5	Fluent, natural, and idiomatic Dutch; well-structured and appropriate medical register. Ex - "Ik neem ibuprofen en probeer rust te nemen. Maar het helpt niet echt."

Table 17

Relevance
Points	Definition and Example
1-2	Major topics missing (symptoms, medication use, lifestyle, and laboratory results.), many irrelevant details, or missing clinical actions. Ex - "Patiënt: Ik houd van voetbal en mijn huisdier is een kat." (Unrelated information to context and/or of topics being discussed.)
3	Some or all target topics touched but detail may be shallow, or topic coverage uneven; some off-topic content. Lacks full coverage. Ex - "Arts: Wat betreft beweging, misschien kunt u een beetje meer wandelen."
4-5	Every target topic is addressed with sufficient detail; all info relevant to consultation. Ex - "Arts: We kunnen naar fysiotherapie of yoga kijken. Ook uw labresultaten wijzen op een geleidelijk probleem." (Addresses symptoms, medication, lifestyle, and lab results directly.)

Table 18

Clinical Use
Points	Definition and Example
1-2	Dialogue is unrealistic for actual clinical setting (e.g., unsafe advice, fundamental errors, implausible behaviors). Ex - unsafe or implausible advice is given ex - "Arts: Neem gewoon meer pijnstillers, zoveel als u wilt, en de rest is niet belangrijk."
3	Could work in clinic with edits; mostly safe and realistic, but has notable weaknesses. Ex - "Arts: Probeer rustig te blijven en misschien wat meer water drinken." (Safe but incomplete; lacks specific clinical recommendation or next steps.)
4-5	Realistic, safe, plausible for clinical scenario; could be useful for annotation/training. Ex - Arts: We zullen uw bloeddruk monitoren en indien nodig medicatie voorschrijven. Houd een dagboek bij van uw symptomen en neem contact op als ze verergeren."

Table 19

Synthetic Data Generation for Language Models

Synthetic data generation for language models creates artificial text, tasks, and interaction traces to train or evaluate models.

See all 292 papers

Decoupling Patient Personas and Clinical Workflows: A Dual-Agent Approach to Realistic Medical Dialogue Generation
Oct 23, 2026Qiyuan Li, Xiaomei Guo, Chao Gao +4Synthetic Clinical Data Generation Text Generation
School of Computer Science and Technology, Wuhan University of Science and Technology, China · Department of Cardiology, Tongji Hospital, Huazhong University of Science and Technology, China · Department of Computer Science, Vrije Universiteit Amsterdam, Netherlands +1
SynDocDis: A Metadata-Driven Framework for Generating Synthetic Physician Discussions Using Large Language Models
Mar 16, 2026Beny Rubinstein, Sergio MatosSynthetic Data Generation Synthetic Data Generation for Language Models
University of Aveiro, Portugal · IEETA, DETI, LASI, University of Aveiro, Portugal
MedAidDialog: A Multilingual Multi-Turn Medical Dialogue Dataset for Accessible Healthcare
Mar 25, 2026Shubham Kumar Nigam, Suparnojit Sarkar, Piyush PatelMedical LLMs Text Generation
University of Birmingham, Dubai, United Arab Emirates · Heritage Institute of Technology, Kolkata, India · Madan Mohan Malaviya University of Technology, India

Oct 23, 2026J/K move · Enter open · S save

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

Qiyuan Li, Xiaomei Guo, Chao Gao, Haijiang Liu+3

School of Computer Science and Technology, Wuhan University of Science and Technology, China · Department of Cardiology, Tongji Hospital, Huazhong University of Science and Technology, China · Department of Computer Science, Vrije Universiteit Amsterdam, Netherlands+1

Synthetic Clinical Data Generation Text Generation Large Language Model-Based Role-Play Simulation Task-Oriented Dialogue Systems Persona-Conditioned Generation

Open paper AskPDF arXiv

Pre-consultation medical dialogues are essential for collecting comprehensive patient information, but are costly to obtain at scale due to privacy and ethical constraints. Existing LLM-based synthesis methods often rely on prompt engineering or rigid rules, limiting their ability to balance clinical relevance with realistic patient diversity. We propose CoRPS-Dial, a controllable dual-agent framework for pre-consultation medical dialogue synthesis that integrates patient personas with guideline-aligned clinical workflow rules. CoRPS-Dial generates dialogues through interactions between a persona-driven patient agent and a process-guided doctor agent, followed by post-hoc quality control for role consistency, workflow completeness, and textual consistency. Experiments on coronary heart disease demonstrate that CoRPS-Dial improves alignment between synthetic and real dialogues in information coverage, clinical compliance, and dialogue naturalness. Downstream fine-tuning experiments further confirm the utility of the generated synthetic data for medical dialogue tasks. These results highlight the effectiveness of CoRPS-Dial for clinically grounded and controllable medical dialogue synthesis.

Abstract

Synthetic Data Generation for Language Models

Explore similar work

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

SynDocDis: A Metadata-Driven Framework for Generating Synthetic Physician Discussions Using Large Language Models

MedAidDialog: A Multilingual Multi-Turn Medical Dialogue Dataset for Accessible Healthcare