SimPatient

Scaffolding Empathy: Training Counselors with Simulated Patients and Utterance-level Performance Visualizations

Venue: CHI 2025
Paper: ACM DL

Overview

SimPatient models patients dealing with alcohol misuse using a cognitive model that tracks internal states. The patient's responses evolve based on the therapist's approach, with a unique feature of generating between-session events that reflect real-world patient experiences between appointments.

Key Features

• Cognitive Model: Tracks 4 internal states (control, efficacy, awareness, reward) on 1-10 scales
• Dynamic State Updates: Internal states evolve based on therapist interactions
• Between-Session Events: Generates realistic events that occur between therapy sessions
• Session Continuity: Can continue from previous session state and history
• Motivational Interviewing Focus: Designed for MI-based counseling training

How It Works

1. Profile Loading: Loads patient persona and initial cognitive model state
2. Session Initialization: Optionally continues from previous session with updated states
3. Response Generation: Generates responses based on current cognitive model
4. State Update: After each interaction, updates the four cognitive dimensions
5. Between-Session Generation: Creates realistic events for multi-session continuity

Cognitive Model

SimPatient tracks four dimensions of patient psychology:

Dimension	Description	Scale
patient_control	Perceived control over drinking behavior	1-10
patient_efficacy	Self-efficacy for making changes	1-10
patient_awareness	Awareness of the problem and its consequences	1-10
patient_reward	Perceived reward/benefit from current behavior	1-10

Usage

CLI

# New session
uv run python -m examples.simulate client=simPatient therapist=MI

# Continue from previous session
uv run python -m examples.simulate client=simPatient client.continue_last_session=true

Python

from omegaconf import OmegaConf
from patienthub.clients import get_client

config = OmegaConf.create({
    'agent_type': 'simPatient',
    'model_type': 'OPENAI',
    'model_name': 'gpt-4o',
    'temperature': 0.7,
    'max_tokens': 1024,
    'max_retries': 3,
    'data_path': 'data/characters/SimPatient.json',
    'data_idx': 0,
    'continue_last_session': False,
    'conv_history_path': 'data/sessions/SimPatient/session_1.json',
})

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

response = client.generate_response("How have things been going since we last talked?")
print(response)

Configuration

Option	Description	Default
data_path	Path to character file	data/characters/SimPatient.json
data_idx	Character index	0
continue_last_session	Resume from previous session	False
conv_history_path	Path to previous session history	data/sessions/SimPatient/session_1.json

Character Data Format

{
  "persona": {
    "age": 34,
    "gender": "male",
    "ethnicity": "Asian",
    "occupation": "software engineer",
    "mbti": "INTJ-A"
  },
  "cognitive_model": {
    "patient_control": 3,
    "patient_efficacy": 4,
    "patient_awareness": 7,
    "patient_reward": 8
  },
  "between_session_event": "After a stressful day at work, the patient received a call from a college friend inviting him to a small get-together..."
}

Between-Session Events

SimPatient can generate realistic between-session events:

# The system automatically generates events like:
"After a stressful day at work, the patient received a call from
a college friend inviting him to a small get-together. He hesitated
but decided to go, thinking he could handle it. However, once there,
the sight and smell of alcohol triggered his cravings..."

State Update Example

After a therapist message about coping strategies:

# Before interaction
cognitive_model = {
    "patient_control": 3,
    "patient_efficacy": 4,
    "patient_awareness": 7,
    "patient_reward": 8
}

# After positive therapeutic interaction
cognitive_model = {
    "patient_control": 4,  # Slight improvement
    "patient_efficacy": 5, # Increased confidence
    "patient_awareness": 8, # Better understanding
    "patient_reward": 7    # Reduced perceived reward from drinking
}