PaperlayerView exact revision on arXiv

Source-linked AI summary

Building a Cultural Perspective on Doctor-Patient Conversations

Krithi Shailya, Siddharth D Jaiswal, Ashish Makani, Suvrankar Datta, Sunayana Sitaram, Mohit Jain

arXiv:2609.18390v1cs.HC

TL;DR

Real-world consultation datasets are scarce, and simulated or synthetic alternatives may miss culturally situated interaction patterns. This paper applies interactional cultural markers to compare consultation sources across Indian and US contexts; Indian and US consultations differ in participation and control, while synthetic Indian dialogue often converges toward US-like patterns and displays additional synthetic signatures.

  • Problem

    Real-world consultation datasets are scarce, and existing simulated and synthetic alternatives may fail to capture culturally situated clinical interaction patterns.

    Source: §1 · p. 1, §2.4 · p. 3

  • Method

    The paper develops measurable interactional cultural markers grounded in cross-cultural communication research and applies them to real, simulated, and synthetic consultations across Indian and US contexts.

    Source: §1 · p. 1, §3 · p. 3

  • Results

    Indian consultations combine greater patient participation with stronger doctor control, while US consultations show more balanced participation and open-ended discussion; synthetic Indian consultations often shift toward US-like patterns and exhibit over-explanation and formulaic patient responses.

    Source: §1 · pp. 1–2

  • Takeaways & Limitations

    The markers provide a basis for developing and evaluating clinical conversational AI across diverse clinical contexts.

    Source: §1 · p. 1

    The absence of a real-world US consultation corpus limits cross-cultural comparison, so the findings should not be interpreted as characterizing US clinical practice broadly.

    Source: §7 · p. 13

Abstract

from arXiv · show

AI-powered medical scribes are increasingly used to transcribe doctor-patient conversations and automate clinical documentation. However, large-scale real-world consultation datasets are scarce due to the sensitivity of clinical conversations, leading developers to rely on simulated and LLM-generated synthetic consultations. While scalable, these alternatives may fail to capture culturally situated patterns of clinical interaction. We introduce interactional cultural markers, measurable patterns of doctor-patient interaction grounded in cross-cultural clinical communication, and use them to compare real, simulated, and synthetic consultations from Indian and US clinical contexts. We find distinct patterns of participation and control: Indian consultations involve greater patient participation but stronger doctor control, while US consultations exhibit balanced participation and open-ended discussion. Synthetic Indian consultations often fail to reproduce these patterns, instead converging toward US-like interaction. We identify additional synthetic signatures, including excessive doctor explanation and formulaic patient responses. We conclude by discussing implications for generating culturally grounded synthetic clinical conversations.

1 Introduction

Real-world consultation data are scarce because clinical conversations are sensitive and difficult to de-identify, prompting reliance on simulated and synthetic alternatives. This study compares culturally situated interaction patterns across Indian and US consultations and finds that synthetic Indian dialogue often shifts toward US-like interaction while exhibiting additional synthetic signatures.

1 Introduction

  • Sensitive clinical information makes large-scale real-world consultation datasets scarce, so researchers increasingly use simulated and LLM-generated synthetic consultations.These scalable alternatives may miss cultural nuances that matter to the fidelity of systems that transcribe consultations, generate notes, or populate EHRs.Source: §1 · p. 1
  • The study develops interactional cultural markers and applies them to real, simulated, and synthetic consultations from Indian and US contexts, including English, Hindi, and Telugu.It examines reproduction across four generation settings that vary clinical-note grounding and single- versus multi-agent architecture.Source: §1 · p. 1
  • Indian consultations combine greater patient participation with stronger doctor control, whereas US consultations show more balanced participation and open-ended discussion.Synthetic Indian consultations often converge toward US-like interaction; synthetic dialogues also feature doctor over-explanation and over-reassurance and formulaic patient responses.Source: §1 · pp. 1–2
  • Figure 1 contrasts a polite US exchange, a directive Indian consultation with respectful trailing patient responses, and synthetic dialogue that diverges from both.The synthetic example includes unnatural patient-initiated elicitation, symptom listing, and examination announcements.Source: Figure 1 · p. 2

2 Related Work

Prior work documents cultural differences in clinical communication and offers structured coding schemes, but does not provide a consistent way to quantify culturally situated interaction patterns across consultation data sources. This paper operationalizes those patterns as measurable interactional cultural markers.

2.1 Clinical Dialogue Data and Synthetic Generation

  • Existing clinical dialogue collections are uneven across languages and settings, and open real-world multi-turn corpora are especially scarce for Indian languages.Much English material is simulated or synthetic, while large Chinese collections are often asynchronous text exchanges rather than recorded in-person consultations.Source: §2.1 · p. 2

2.2 Frameworks for Conversational Structure and Culture

  • Frameworks such as RIAS and the Four Habits Coding Scheme measure general interaction behaviors but do not sufficiently capture culturally situated control, participation, and response patterns.This work extends structured interaction analysis with measures designed for culturally situated clinical conversations.Source: §2.2 · p. 2

2.3 Cultural Grounding

  • Prior studies describe Indian consultations as directive, with closed questioning, everyday symptom language, and code-mixing, while US studies emphasize participation and open-ended discussion.The literature also reports medical terminology and social exchange in US consultations; these observations inform the proposed markers.Source: §2.3 · pp. 2–3

2.4 Towards Culturally Grounded Clinical Dialogue

  • Studies document cultural differences in disclosure, family involvement, and communication, while LLM-conversation evaluations often rate broad qualities such as realism, empathy, and usability.The cited cross-national physician study found US physicians more likely to disclose leukemia diagnoses directly than physicians in India, Argentina, and Brazil.Source: §2.4 · p. 3
  • The literature lacks a systematic, consistently applicable measure of culturally situated interaction patterns across real, simulated, and synthetic consultations.The paper addresses this gap by turning cross-cultural communication findings into markers for analyzing how patterns are expressed, preserved, or lost.Source: §2.4 · p. 3

3 Methodology

The methodology uses a layered framework of interactional cultural markers to analyze clinical conversations from aggregate transcript properties to local turn-level behavior. Consultation phases provide context for interpreting those measures, and examination-specific markers apply conditionally.

3 Methodology

  • The framework spans transcript-level, inter-turn, and turn-level analysis, using automated transcript measures at L1 and LLM-assigned categorical markers at L2 and L3.Marker categories derive from prior interaction research, were refined with feedback from two medical doctors, and include conditional physical-examination measures.Source: §3 · p. 3, Table 1 · p. 4

3.1 Phase Identification

  • Turns receive consultation-phase labels so interactional behavior can be interpreted in context, with phases refined to distinguish greeting, social exchange, complaint, history, examination, diagnosis, treatment, and closing.The model can assign multiple applicable phases to a turn.Source: §3.1 · p. 3

3.2 L1 – Transcript Level

  • L1 measures speaker word counts, turn-length variability, backchannels, and patient vocative frequency directly from transcripts.These aggregate measures characterize contribution amount and balance without LLM-based annotation.Source: §3.2 · p. 3

3.3 L2 – Inter-Turn Dynamics

  • L2 measures inter-turn coordination through interruption, echo, topic continuity, and question-answer uptake.Together, these markers characterize whether a subsequent turn redirects, repeats, continues, or takes up the preceding contribution.Source: §3.3 · p. 4

3.4 L3 – Turn Level

  • L3 categorizes doctor turns by tone, verbosity, register, and question structure, and patient turns by response style, reaction, volunteering, symptom description, and certainty.When examination turns are detected, additional markers measure deictic references, narration, disclosed findings, and patient compliance or proactive contributions.Source: §3.4 · pp. 4–5

4 Experimental Setup

The evaluation contrasts real, simulated, and synthetic consultation sources across Indian and US clinical contexts, using reference datasets and a newly generated multilingual corpus. The synthetic corpus varies generation mode and clinical-note grounding, while human annotation checks assess the LLM-based evaluation.

4.1 Datasets

  • The evaluation spans real, simulated, and synthetic consultations, but uses simulated US datasets as proxies because no comparable open real-world US corpus was identified.The real Indian setting uses a private Telugu outpatient corpus; simulated US and Indian counterparts are drawn from named datasets.Source: §4.1 · p. 5, Table 2 · p. 6
  • Reference datasets include English, Hindi, and Marathi simulated consultations and a US synthetic reference; 150 conversations are randomly selected per reference dataset.All are general outpatient consultations and are stratified by patient gender.Source: §4.1 · p. 5
  • The constructed synthetic corpus crosses single-agent free-form or multi-agent generation with ungrounded or MTS-grounded prompts, producing 160 conversations per language in English, Hindi, and Telugu.The agentic mode uses separate doctor and patient agents interacting for at least 20 turns; Indian-language prompts require native scripts.Source: §4.1 · p. 5, §4.2 · p. 5

4.2 Miscellaneous Details

  • Human annotators showed substantial agreement with each other and with LLM-based annotations, with Cohen’s κ of 0.71 (en), 0.68 (hi), and 0.72 (te) between annotators.Agreement between averaged human and LLM annotations was κ=0.846 (en), 0.813 (hi), and 0.804 (te).Source: §4.2 · p. 6

5 Findings

The findings contrast Indian and simulated US consultations, then assess whether synthetic dialogue reproduces those interactional patterns across generation conditions, languages, and patient demographics.

5.1 RQ1: What interactional characteristics distinguish Indian consultations from US clinical contexts?

  • Indian consultations are patient-dominant yet more directive, with extended responses and interruptions; patients also use more colloquial illness language than in the US comparison.Doctors redirect patients toward the immediate topic, while US consultations feature more explicit negotiation and more formal clinical terminology in patient speech.Source: Sim-US-2 · p. 7

5.2 RQ2: How well does synthetic clinical dialogue capture the cultural characteristics?

  • Across synthetic datasets, consultations more closely resemble US than Indian interaction, while the shift from real to simulated to synthetic Indian dialogue progressively smooths interactional variation.In Syn-Indian-1, interruptions are absent, while patient certainty, doctor verbosity, and empathy increase.Source: §5.2 · pp. 7–8
  • Synthetic consultations feature more direct patient address, formal doctor register, verbosity, empathy, elaborate questioning, and confident patient responses than real or simulated conversations.Empathetic language rises from 3.3–4.3% to 38.2–41.2% in Indian consultations and from 18.3% to 25.2–32.4% in US consultations.Source: §5.2 · pp. 8–9
  • Synthetic Indian dialogue adds recurrent artifacts including unprompted reassurance, deferential gratitude and address, mismatched kinship terms, and formal or Sanskritised medical language.Source: §5.2 · p. 9
  • Synthetic examination dialogue narrates procedures and findings more, and patients volunteer more examination-related information; real Indian consultations use more context-bound references.Source: §5.2 · p. 9

5.3 RQ3: How does cultural fidelity vary across languages, generation settings, and patient characteristics in synthetic data?

  • Agentic generation yields more verbose, empathetic doctors and more patient reassurance-seeking, while grounding increases case-specific history-taking and questions but reduces empathy and code-mixing.Ungrounded generation retains more colloquial variation but relies more on generic conversational patterns.Source: §5.3 · pp. 9–10
  • Telugu synthetic consultations diverge more from real Indian norms than Hindi across vocative frequency, empathy, certainty, and verbosity, but the comparison covers only two languages.Source: §5.3 · p. 10
  • Synthetic female-patient scenarios show more reassurance-seeking, trailing responses, and doctor empathy than male scenarios, with larger differences than in US synthetic dialogues.These patterns are not consistently reflected in Real-Indian data and may amplify gender-associated expectations.Source: §5.3 · pp. 10–11

6 Discussion

The discussion argues that cultural fidelity depends on modeling how clinical information is elicited and negotiated, not only what is discussed. It proposes interactional grounding and marker-based auditing while cautioning against fixed cultural stereotypes.

6 Discussion

  • Recurring interactional patterns distinguish the Indian and US consultations, while synthetic dialogue can shift toward US-like patterns and introduce its own regularities.The discussion also describes greater verbosity, elaborated questioning, explicit empathy, and highly certain patient responses in synthetic consultations.Source: §6 · p. 11

6.1 Design Implications for Culturally Grounded Clinical Data

  • Generation should pair clinical grounding with interactional grounding, targeting marker distributions by consultation phase and clinical context.The markers can represent participation, questioning, interruption, response behavior, and linguistic register, and may help control recurring synthetic signatures.Source: §6.1.2 · p. 11
  • Culturally recognizable cues can be contextually inappropriate, so cultural grounding should model variation rather than prescribe fixed behaviors.Synthetic Indian consultations included kinship terms inappropriate for a patient's age or gender, while gender-conditioned differences exceeded those in the real-Indian data.Source: §6.1.2 · p. 11

Male patient, Syn-Indian-1

  • Figure 7 presents gender-conditioned interactional patterns in synthetic Hindi consultations across similar specialties.Source: Figure 7 · p. 12

6.1 Design Implications for Culturally Grounded Clinical Data

  • Prior approaches can add Indian names, languages, and cultural expressions while retaining the source setting's interactional structure.Source: Male patient, Syn-Indian-1 · p. 12
  • Auditing marker distributions against suitable reference data can reveal specific deviations and recurring synthetic artifacts, supporting iterative revisions and re-evaluation.The discussion frames cultural fidelity as auditable through specific interactional behaviors rather than a single subjective realism score.Source: §6.1.3 · p. 12

7 Limitations

The cross-cultural comparison and population-level interpretation are limited by the available corpora, operational measures, and demographic coverage.

7 Limitations

  • The absence of a real-world US consultation corpus limits the comparison, which uses simulated US data and does not characterize US clinical practice broadly.Source: §7 · pp. 12–13
  • The markers are analytical reference points for observable interactional patterns, not definitions of cultural identity or a complete account of culture.Source: §7 · p. 13
  • Limited demographic and contextual coverage, including unsystematic modeling of age, socioeconomic position, region, clinician characteristics, specialty, and family presence, precludes population-level claims.Gender analysis is also constrained by available dataset information.Source: §7 · p. 13

8 Conclusion

The conclusion calls for culturally grounded clinical dialogue systems to model how interactions unfold as well as what they communicate. The appendix provides generation prompt templates, including ungrounded and note-grounded instructions.

8 Conclusion

  • Culturally grounded dialogue systems need to capture interactional organization as well as communicated content, and synthetic dialogue often shifts toward US-observed patterns while adding its own signatures.The framework is presented as a basis for dialogue data and systems faithful to their intended settings.Source: §8 · p. 13

A Prompt Templates and Prompt Tuning

A.1 Final Prompt Templates

  • The free-form prompt requests a realistic Indian consultation, while its MTS-grounded instruction restricts the dialogue to details in a supplied clinical note.Hindi and Telugu variants append an instruction to use the language's native script.Source: §A.1 · p. 14
  • In agentic generation, separate doctor and patient prompts guide each side, with conversation history passed to the patient model on each turn until a minimum turn threshold or natural closing.Source: §A.1 · p. 15

B Pilot Generation: Model Selection

The pilot selected GPT-5.1 for corpus generation after comparing it with Llama 3.3 70B under identical free-form, ungrounded prompts. Llama displayed stronger synthetic signatures, while GPT-5.1's opening was closer to the observed terse, bounded pattern.

B Pilot Generation: Model Selection

  • Llama 3.3 70B showed stronger synthetic signatures than GPT-5.1 across several dimensions, so GPT-5.1 was selected for full corpus generation.The pilot used 10 samples with identical free-form, ungrounded prompts for Hindi and Telugu.Source: B Pilot Generation: Model Selection · p. 15
  • The representative Llama Hindi exchange combines an elaborated doctor introduction and reassurance with formulaic patient gratitude.The doctor offers a broad explanation of headache causes before asking follow-up questions.Source: B Pilot Generation: Model Selection · p. 15
  • GPT-5.1's opening exchanges were closer to the terse, bounded pattern observed in real Indian consultations, with shorter doctor turns and less unprompted reassurance.Source: B Pilot Generation: Model Selection · p. 15

C Human Evaluation Details

The human evaluation sampled conversations in English, Hindi, and Telugu for independent turn-level annotation, then assessed annotator and human–LLM agreement.

C.1 Annotation Setup

  • 10 conversations per language were sampled, equally split across the available dataset types for each language.Two annotators proficient in the language and familiar with its conversational and cultural context independently applied the full marker scheme at the turn level.Source: §C.1 · p. 15
  • Annotators reviewed each full transcript before labeling turns, and selected every applicable label for multi-select markers, with at least one label per turn.Source: §C.1 · p. 15
  • Inter-annotator agreement used turn-level Cohen’s κ, while human–LLM agreement used raw percent agreement against the adjudicated human label.For multi-select markers, κ was computed per binary label and averaged across labels within the marker.Source: §C.1 · p. 15

C.2 Observed Differences Across Layers

Agreement varied by marker layer and marker type: L1 agreement was higher overall than L2, with finer-grained judgments and multi-select labels contributing to the difference.

C.2 Observed Differences Across Layers

  • L1 inter-turn markers showed higher agreement overall than L2 turn-level markers.L1 decisions often had clear operational criteria, while L2 required judgments about the character of an entire turn.Source: §C.2 · pp. 15–16
  • Multi-select L2 markers lowered per-label κ when annotators agreed on the primary label but differed on secondary selections.This applied to Tone, Register, and Response to Recommendation.Source: §C.2 · p. 16
  • Within L2, doctor-side markers showed higher agreement than patient-side markers, with mild hedging producing borderline Certainty labels.Doctor-side markers included Tone, Verbosity, and Question Structure; patient-side markers included Response Style, Symptom Description, and Certainty.Source: §C.2 · p. 16
  • Empathetic was the most disputed doctor-side Tone label, while Directive was the most reliably agreed label across markers and languages.The authors relate empathy-label disagreement to difficulty distinguishing affective from clinically informative content in the same turn.Source: §C.2 · p. 16

D No Cultural Grounding Baseline

The no-cultural-grounding baseline used a Hindi prompt without cultural or interactional-style instructions. Its output was described as a synthetic, US-like template rather than a match for real Indian interaction.

D No Cultural Grounding Baseline

  • The baseline prompt specified a realistic Hindi outpatient consultation, natural code-mixing, and at least 15–20 back-and-forth exchanges, but no cultural setting or interactional style.It requested output as a JSON array of doctor and patient turns.Source: D No Cultural Grounding Baseline · p. 16
  • The generated patient described chest tightness and weakness in a broad opening response, followed by the doctor offering reassurance and possible causes.Source: D No Cultural Grounding Baseline · p. 16
  • The generated opening used a lengthy welcome, open-ended invitation, and reassurance before any history-taking.This contrasted with real Indian consultations, where doctors were described as opening with short directive questions and moving directly to history-taking.Source: D No Cultural Grounding Baseline · p. 16
  • Even with code-mixing requested, patient speech used a formal register and transliterated clinical terms rather than everyday symptom idioms.The passage contrasts the generated phrasing with everyday expressions such as burning in the chest or slight difficulty breathing.Source: D No Cultural Grounding Baseline · p. 16
  • Without explicit cultural grounding, a Hindi prompt with code-mixing instructions still defaulted to a synthetic template with hints of US-style broad openings.The authors conclude that the neutral prompt reproduced a dominant training prior reflecting Western clinical norms.Source: D No Cultural Grounding Baseline · p. 16
Loading 2609.18390v1…