PaperlayerView exact revision on arXiv

Source-linked AI summary

Modeling Challenging Patient Interactions: LLMs for Medical Communication Training

Anna Bodonhelyi, Christian Stegemann-Philipps, Alessandra Sonanini, Lea Herschbach, Marton Szep, Anne Herrmann-Werner, Teresa Festl-Wietek, Enkelejda Kasneci, Friederike Holderried

arXiv:2503.22250v2cs.HCcs.AI

TL;DR

LLM virtual patients often lack diverse, emotionally charged communication styles despite their importance in medical training. This study models Satir’s accuser and rationalizer styles with structured prompts, finding that participants generally judged the interactions authentic and recognized intended styles, though the accuser was sometimes misclassified.

  • Problem

    Many LLM virtual patients prioritize clinical accuracy or dialogue coherence while overlooking emotionally charged communication styles important to realistic patient interactions.

    Source: §1 · pp. 2–3

  • Method

    The study used targeted prompts, author’s notes, and first messages to build medically grounded virtual patients embodying Satir’s accuser and rationalizer styles.

    Source: §5.1 · p. 13, §6 · p. 25

  • Results

    Most participants found the conversations authentic and realistic and selected style-specific adjectives, although the accuser was often misclassified as congruent or rationalizer.

    Source: §6 · p. 26, §7 · p. 27

  • Takeaways & Limitations

    The findings support the potential of communication-style virtual patients for realistic medical-training simulations and assessment.

    Source: §7 · p. 27

    The professional conference participant pool restricted input from other groups, including students, whose perspectives could broaden usability evaluation.

    Source: §6 · p. 25

Abstract

from arXiv · show

Effective patient communication is pivotal in healthcare, yet traditional medical training often lacks exposure to diverse, challenging interpersonal dynamics. To bridge this gap, this study proposes the use of Large Language Models (LLMs) to simulate authentic patient communication styles, specifically the "accuser" and "rationalizer" personas derived from the Satir model, while also ensuring multilingual applicability to accommodate diverse cultural contexts and enhance accessibility for medical professionals. Leveraging advanced prompt engineering, including behavioral prompts, author's notes, and stubbornness mechanisms, we developed virtual patients (VPs) that embody nuanced emotional and conversational traits. Medical professionals evaluated these VPs, rating their authenticity (accuser: $3.8 \pm 1.0$; rationalizer: $3.7 \pm 0.8$ on a 5-point Likert scale (from one to five)) and correctly identifying their styles. Emotion analysis revealed distinct profiles: the accuser exhibited pain, anger, and distress, while the rationalizer displayed contemplation and calmness, aligning with predefined, detailed patient description including medical history. Sentiment scores (on a scale from zero to nine) further validated these differences in the communication styles, with the accuser adopting negative ($3.1 \pm 0.6$) and the rationalizer more neutral ($4.0 \pm 0.4$) tone. These results underscore LLMs' capability to replicate complex communication styles, offering transformative potential for medical education. This approach equips trainees to navigate challenging clinical scenarios by providing realistic, adaptable patient interactions, enhancing empathy and diagnostic acumen. Our findings advocate for AI-driven tools as scalable, cost-effective solutions to cultivate nuanced communication skills, setting a foundation for future innovations in healthcare training.

1. Introduction

Existing virtual patients often prioritize clinical accuracy or coherent dialogue over diverse, emotionally charged communication styles. This study examines whether LLM-powered virtual patients can simulate two challenging styles while remaining medically accurate.

1. Introduction

  • Existing virtual patients often overlook emotionally charged communication styles needed to prepare trainees for real-world interpersonal dynamics.Many systems prioritize clinical accuracy or general dialogue coherence.Source: §1 · pp. 2–3
  • The study evaluates LLM-powered virtual patients modeling the Satir accuser and rationalizer styles for realism and alignment with intended behavior.It also explores emotion and sentiment analysis as quality-control tools.Source: §1 · p. 3
  • The research questions address prompt techniques for distinct styles, professional ratings of realism and style adherence, and alignment with emotion and sentiment measures.The styles are defined by the Satir model.Source: §1 · p. 3

2. Related Work

The related work frames challenging patient communication styles as a target for conversational AI in medical education, where realistic, emotionally varied interactions remain an open need.

2. Related Work

  • The review covers patient communication styles, conversational-AI modeling, and chatbot applications in medical education.Source: §2 · pp. 3–4

2.1. Patient Communication Styles

  • The Satir model describes five stress-related styles; this study focuses on the challenging accuser and rationalizer types.The accuser is confrontational, while the rationalizer adopts a detached, logical stance.Source: §2.1 · p. 4

2.2. Modeling Communication Styles in Conversational AI

  • Unlike general style models, Satir’s archetypes connect speech and content to stress responses; this work prioritizes authentic personality representation over engagement.The study uses prompting as a flexible, data-efficient alternative to costly fine-tuning (Petroni et al., 2019; Sz´ep et al., 2024).Source: §2.2 · pp. 5–6

2.3. Chatbots in Medical Education

  • Virtual-patient tools support repeatable, cost-effective practice, but authentic portrayal of distinct challenging personalities remains uncertain, especially under stress or emotional distress.Existing tools can deliver plausible or script-aligned responses, yet may have accuracy or evaluation challenges.Source: §2.3 · pp. 6–7

3. Methodology

The study builds accuser and rationalizer virtual patients on an existing LLM role-play framework, adapting it with Satir-style behaviors and structured prompts for realistic, challenging interactions.

3. Methodology

  • The study develops accuser and rationalizer virtual patients based on the Satir model within an existing clinical-skills framework.Source: §3 · p. 8, §3.1 · p. 8

3.1. Framework

  • The framework was adapted from history-taking feedback to a participant questionnaire and supports English- and German-language interactions.Source: §3.1 · p. 8

3.2. Technical details

  • The framework adds Satir-derived interaction patterns tailored to challenging scenarios, extending the original tool’s behavioral diversity.Source: §3.2 · p. 8

3.3. Prompt Engineering

  • Behavioral prompts, author’s notes, and first messages reinforce each patient’s style, while structured illness-script fields encode patient history and communication tendencies.The author’s note continuously reinforces personality, tone, and behavior; stubbornness instructions guide resistance to psychological explanations and therapy.Source: §3.3 · pp. 9–10

4. Procedure of the User Study

The user study gathered psychotherapy professionals’ interactions with virtual patients under a controlled procedure, then assessed participant feedback alongside emotion and sentiment measures.

4. Procedure of the User Study

  • Participants used a fixed study environment, chose English or German, were randomly assigned a virtual patient, and held an open-ended conversation before completing a questionnaire.The questionnaire focused on communication-style consistency and educational value.Source: §4 · p. 10

4.2. Participant Recruitment

  • The study recruited 15 participants for the accuser virtual patient and 17 for the rationalizer at an international psychotherapy conference.Participants could withdraw, and responses were collected anonymously with consent.Source: §4.2 · p. 11

4.3. Analysis

  • The evaluation framework used a 5-point Likert scale to assess VP interactions, including character genuineness, with participants evaluating the authenticity of VP responses.Table 2 presents demographic information about participants interacting with the accuser and rationalizer VPs.Source: §4.3 · p. 11, Table 2 · p. 13
  • Emotion analysis summarized predicted scores for 53 emotions across conversations, while sentiment analysis used the highest-scoring dimension on a nine-level scale.The analyses were conducted separately for each virtual patient’s communication style.Source: §4.3 · pp. 12–13

5. Results

The results show that prompt design helped LLM virtual patients sustain distinct communication styles and that participants generally rated the accuser and rationalizer as authentic and realistic. Emotion and sentiment analyses also produced profiles consistent with the intended personas.

5.1. Communication Style Development

  • The approach used behavioral instructions, predefined first messages, author’s notes, and style-specific constraints to maintain communication consistency and align VP responses with intended Satir archetypes.On five selected user-specific question-answer pairs rated on a 5-point Likert scale, realism averaged 3.7±0.8 for the accuser and 3.5±1.1 for the rationalizer; matching adjectives were chosen in 56.4% and 41.5% of cases, respectively.Source: §5.1 · p. 13, §5.3 · pp. 20–21

5.1.1. Author’s Note and First Message

  • The first message and author’s note established tone and emotional context; without them, VPs tended toward generic neutrality and inconsistent role adherence.The author’s note carried the full illness script and was dynamically inserted into the chat history to support coherence.Source: §5.1.1 · pp. 14–15

5.1.2. Behavior Instruction Prompts

  • Behavioral instructions in case information reinforced style consistency, but VPs initially drifted toward neutral tone after 4–6 prompts.More detailed mood- and situation-related instructions improved sustained representation of the intended styles.Source: §5.1.2 · pp. 15–16

5.1.3. Stubbornness Mechanism

  • The stubbornness mechanism allowed both VPs to accept psychotherapy after users addressed emotions and connected symptoms with life events.Six participants addressed the accuser’s emotions and assessed symptoms, five did so for the rationalizer, and both VPs eventually accepted psychotherapy suggestions.Source: §5.1.3 · pp. 16–17

5.1.4. Lessons from Unsuccessful Prompting Strategies

  • Predefined answers strengthened stylistic reliability but produced repetitive, rigid replies to similar health-related follow-up questions, so this strategy was excluded from the final case information.For less-defined topics unrelated to symptoms or medical history, predefined options could enhance stylistic accuracy and consistency.Source: §5.1.4 · pp. 17–18

5.2. AI Familiarity and Engagement among Participants

  • Participants showed a cautious but optimistic outlook toward AI, with notable trust and reported potential to enhance daily routines.The study also asked about personal use of AI applications such as ChatGPT.Source: §5.2 · p. 19

5.3. Authenticity and Realism of Conversational Agent

  • On a 5-point Likert scale, participants rated authenticity 3.8±1.0 for the accuser and 3.7±0.8 for the rationalizer.Participants also identified the intended styles: a notable portion selected accuser and a majority selected rationalizer among respondents choosing a Satir style (Figure 2).Source: §5.3 · pp. 19–20, Figure 2 · p. 20

5.4. Illness Scripts and Emotional Alignment

  • Hume AI emotion analysis found the accuser profile aligned with pain, distress, annoyance, and anxiety, while the rationalizer showed contemplation, confusion, calmness, and concentration.The accuser’s sentiment averaged 3.1±0.6 and the rationalizer’s 4.0±0.4 on a scale from one (extremely negative) to nine (extremely positive).Source: §5.4 · pp. 22–25, Table 5 · p. 23, Table 6 · p. 24

6. Discussion and Limitations

The study used targeted prompts to create accuser and rationalizer virtual patients, then evaluated their perceived communication styles. Its discussion notes limits from the participant pool, the two-style design, patient descriptions, and untested LLM differences.

6. Discussion and Limitations

  • Behavioral prompts, author’s notes, and first messages encoded emotional tone, response patterns, and linguistic traits to align VP responses with predefined styles.The study focused on the accuser and rationalizer styles from the Satir model.Source: §6 · p. 25
  • The evaluation drew mainly on psychological-conference professionals, limiting perspectives from students and other user groups; non-native language proficiency and sample size may also have affected assessments.The authors propose more diverse participant pools for broader evaluation.Source: §6 · p. 25
  • Only accuser and rationalizer styles were modeled, and both patients’ stubborn refusal of therapy may have influenced their communication.The authors suggest broader behaviors and conditions, including additional styles or neutral patients, for greater versatility.Source: §6 · pp. 25–26
  • The Satir model’s extreme labels were considered unsuitable for nuanced, adaptable LLM styles, so the illness scripts added emotions beyond those specified by the model.The authors also plan to incorporate contextual information such as patient history or diagnostic results.Source: §6 · pp. 26–27
  • “Overstimulated” was frequently selected for both VPs, which the authors suggest may reflect reinforcement learning from human feedback encouraging responsive, active participation.The authors note a possible bias toward heightened emotional intensity.Source: §6 · pp. 26–27
  • The study did not investigate how LLM type affects VP performance, although the framework is designed to accommodate different LLMs for future evaluation.The authors identify other LLM integrations as a possible route to optimizing accuracy and realism.Source: §6 · p. 27

7. Conclusion

The study concludes that LLM-driven virtual patients modeled with specific communication styles can provide realistic, engaging simulations for medical training and assessment. Participant feedback and automated emotion analysis supported the intended styles.

7. Conclusion

  • Participant feedback and automated emotion analysis supported the intended behaviors of the accuser and rationalizer VPs, which the authors conclude have potential for realistic medical-training interactions.Most participants described the conversations as authentic and realistic and selected style-specific adjectives.Source: §7 · p. 27

Funding

The paper reports that it received no specific grant funding.

Funding

  • The research received no specific grant from funding agencies in the public, commercial, or not-for-profit sectors.Source: Funding · p. 28

Ethical Statement

The authors state that the study complied with relevant laws and institutional guidelines and received institutional approval and participant consent.

Ethical Statement

  • The protocol was approved by the institutional committee on 29.10.2024 under reference 565/2024BO2, and participants gave informed consent.The authors report compliance with relevant laws and institutional guidelines.Source: Ethical Statement · p. 28

Data Statement

The data are available to researchers on request, and Appendix A presents participant employment distribution.

Data Statement

  • Study data are available upon request for research purposes, subject to privacy and confidentiality regulations.Source: Data Statement · p. 28

Appendix A. Employment Distribution

  • Appendix A presents employment distribution for both virtual patients, with Table A.7 identified as the participant employment-distribution table.Source: Appendix A · p. 34, Table A.7 · p. 34

Appendix B. Accuser VP Detailed Description

The appendices describe the accuser and rationalizer virtual patients and provide the questionnaire used to evaluate their simulations.

Appendix B. Accuser VP Detailed Description

  • Gerhard Anton is a 53-year-old bus driver with chronic hip and shoulder pain, frustrated by ineffective treatment and skeptical of psychological explanations.His pain worsened after a bus-driving car accident; he is defensive when he feels dismissed but may cooperate when understood and respected.Source: Appendix B · pp. 34–35

Appendix C. Rationalizer VP Detailed Description

  • Andreas Petersen is a stressed automotive manager with recent weight loss, sleep and concentration problems, and a strong belief that his symptoms have a physical cause.He presents as composed and emotionally distanced, while feeling vulnerable about symptoms affecting his professional self-image.Source: Appendix C · pp. 35–36

Appendix D. Questionnaire

  • The questionnaire asks respondents to rate answer realism, communication-style simulation, educational usefulness, and authenticity, and to identify the Satir style they recognized.Source: Appendix D · pp. 36–38
  • The questionnaire also asks respondents which emotions or mental states the virtual patient conveyed and invites explanations for disagreement or suggested app changes.Source: Appendix D · pp. 36, 37, 40

Appendix E. Hume AI Emotion Prediction

The appendix lists the 53 emotions for which Hume AI predicts scores.

Appendix E. Hume AI Emotion Prediction

  • Hume AI predicts scores for 53 emotions, including anger, calmness, contemplation, pain, and tiredness.Source: Appendix E · p. 40
Loading 2503.22250v2…