AdaptiveVP

A Framework for LLM-Based Virtual Patients that Adapts to Trainees' Dialogue to Facilitate Nurse Communication Training

Venue: ACL 2025 (Findings)
Paper: ACL Anthology

Overview

AdaptiveVP creates virtual patients for nurse communication training that adapt their responses based on the trainee's communication quality. It evaluates nurse responses on multiple dimensions (tone, empathy, de-escalation techniques) and adjusts patient behavior accordingly, making it ideal for training healthcare professionals in handling difficult patient interactions.

Key Features

• Adaptive Responses: Patient behavior adapts based on nurse's communication quality
• Multi-Dimensional Assessment: Evaluates tone (calm, clear), empathy level (0-6 scale), and de-escalation techniques
• Patient Personality Types: Overdependent, Overly Authoritative, Aggressive and Uncooperative
• Stage Directions: Dynamic response guidance based on conversation context
• Rich Output: Generates inner monologue, verbal response, and non-verbal cues

How It Works

1. Nurse Input Analysis: Evaluates the nurse's communication on tone, empathy, and de-escalation
2. Stage Direction Selection: Chooses appropriate response direction based on analysis
3. Response Generation: Creates response with inner thoughts, verbal content, and non-verbal cues
4. Safety Monitoring: Reviews the generated response to ensure it remains within professional boundaries, aligns with the patient profile, and adheres to the adaptation direction.

Patient Types

Type	Description	Typical Behaviors
Dependent	Requires constant attention and reassurance	Frequent calls, anxiety, seeking comfort
Authoritarian	Emphasizes status, demands special treatment	Complaints, threats, dismissive of staff
Aggressive	Responds with hostility when demands not met	Loud voice, threatening, confrontational
Uncooperative	Passive-aggressive or refuses to engage in care	Ignoring staff, feigning sleep, delaying treatment

The original study validated these profiles within a South Korean nursing context, but the framework is adaptable to other cultural settings.

Evaluation Dimensions

AdaptiveVP evaluates nurse communication on:

Tone Assessment

• Calm: Did the nurse suppress contempt, frustration, anger, or anxiety?
• Clear: Did the nurse use clear sentences to reduce confusion?

Empathy Level (0-6 scale)

Assesses the depth of emotional understanding and validation

De-escalation Techniques

• Autonomy: Offering choices, seeking permission
• Limit Setting: Establishing clear behavioral boundaries
• Problem Solving: Clarifying issues, redirecting focus

Prohibited Behaviors

• Premature empathy ("I understand" without justification)
• Invalidating beliefs
• Dismissive commands ("Calm down")

Usage

CLI

uv run python -m examples.simulate client=adaptiveVP therapist=user

Python

from omegaconf import OmegaConf
from patienthub.clients import get_client

config = OmegaConf.create({
    'agent_type': 'adaptiveVP',
    'model_type': 'OPENAI',
    'model_name': 'gpt-4o',
    'temperature': 0.7,
    'max_tokens': 1024,
    'max_retries': 3,
    'data_path': 'data/characters/AdaptiveVP.json',
    'dir_path': 'data/resources/AdaptiveVP_stage_direction.json',
    'data_idx': 0,
})

client = get_client(configs=config, lang='en')
client.set_therapist({'name': 'Nurse'})

response = client.generate_response("I understand you're feeling anxious. Would you like to talk about what's on your mind?")
print(f"Response: {response.content}")
print(f"Inner thoughts: {response.inner_monologue}")
print(f"Non-verbal: {response.non_verbal}")

Configuration

Option	Description	Default
data_path	Path to character file	data/characters/AdaptiveVP.json
dir_path	Path to stage directions	data/resources/AdaptiveVP_stage_direction.json
data_idx	Character index	0

Character Data Format

{
  "id": 0,
  "type": 1,
  "type-text": "Dependent",
  "name": "Im Kyung",
  "situation": "A patient with a dependent tendency continuously calls the nurse...",
  "Chief complaint": "'Nurse, I'm so sorry... My chest is pounding...'",
  "gender": "Female",
  "age": 55,
  "religion": "Non-religious",
  "height": "162cm",
  "weight": "58kg",
  "Main symptom": "Anxiety and insomnia due to breast cancer",
  "History of present illness": "Diagnosed with breast lump 2 weeks ago...",
  "social history": "Freelance designer, divorced, no children",
  "past medical history": "Depression (5 years)",
  "past surgical history & date": "Appendectomy (20 years ago)",
  "family medical history": "Sister - breast cancer",
  "allergies": "Latex",
  "medication": "Escitalopram 10mg daily",
  "primary diagnosis": "Right breast cancer stage 2",
  "communication_summary": "The patient has a dependent personality...",
  "first_statement": "Nurse, could I have a quick word with you?"
}

Response Format

AdaptiveVP generates structured responses with three components:

class Response(BaseModel):
    inner_monologue: str  # Patient's internal thoughts
    content: str          # Verbal response to the nurse
    non_verbal: str       # Non-verbal communication/actions