TalkDep

Clinically Grounded LLM Personas for Conversation-Centric Depression Screening

Venue: CIKM 2025
Paper: ACM DL

Overview

TalkDep creates clinically grounded personas for depression screening research. Each persona has a specific BDI (Beck Depression Inventory) score, depression level, and detailed symptom profiles, making it ideal for training and evaluating depression screening conversational agents.

Key Features

• BDI-II: Strictly constructed based on the Beck Depression Inventory-II (BDI-II).
• BDI-Grounded: By using pre-defined BDI-II scores as a ground truth to evaluate and quantify the performance of the simulation.
• Validated Personas: Constructed 12 validated depression personas, covering four severity levels: Minimal, Mild, Moderate, and Severe.

How It Works

1. Profile Loading: Loads persona with BDI score and symptom profiles
2. System Prompt Construction: Builds prompt incorporating all persona details
3. Conversation Handling: Maintains conversation history for context
4. Response Generation: Generates responses following linguistic patterns and behavioral constraints

Depression Levels

TalkDep personas span the full range of depression severity:

BDI Score	Depression Level	Characteristics
0-9	Minimal Depression	Generally positive, occasional stress
10-18	Mild Depression	Concentration issues, irritability, sleep problems
19-29	Moderate Depression	Persistent sadness, social withdrawal
30-63	Severe Depression	Significant impairment, possible suicidal ideation

Usage

CLI

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

Python

from omegaconf import OmegaConf
from patienthub.clients import get_client

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

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

response = client.generate_response("How have you been feeling lately?")
print(response)

Configuration

Option	Description	Default
data_path	Path to character file	data/characters/talkDep.json
data_idx	Character index	0

Character Data Format

{
  "name": "Alex",
  "age": 24,
  "gender": "Male",
  "bdi_score": 15,
  "depression_level": "Mild Depression",
  "key_negative_symptoms": [
    {
      "name": "Difficulty Concentrating",
      "description": "I find it hard to focus on things, even small tasks.",
      "severity": 1
    },
    {
      "name": "Irritability",
      "description": "I get annoyed easily, even over little things.",
      "severity": 1
    },
    {
      "name": "Sleep Disturbances",
      "description": "I often wake up in the middle of the night.",
      "severity": 2
    }
  ],
  "life_history": [
    "Recently graduated with a degree in Philosophy...",
    "Struggling to find a steady job in his field..."
  ],
  "social_context": [
    "Has a small group of supportive friends...",
    "Uses social media to stay connected but avoids sharing anything personal."
  ],
  "linguistic_patterns": [
    "Often uses phrases like 'I don't know' or 'It's probably nothing'",
    "Appears neutral or slightly dismissive when describing challenges"
  ],
  "emotional_tone": [
    "Generally neutral, with occasional moments of frustration or irritability"
  ],
  "typical_topics": [
    "Struggles with keeping a routine, including sleep and eating habits",
    "Feeling annoyed with himself for not accomplishing as much as he planned"
  ],
  "behavioral_constraints": [
    "Reluctant to openly discuss irritability unless probed",
    "Deflects questions about deeper emotions with neutral responses"
  ],
  "response_goals": [
    "Express frustration about specific challenges while avoiding appearing overly negative"
  ],
  "social_media_activity": [
    "*Example Post:* 'When you spend 20 minutes staring at your to-do list...'",
    "*Typical Interactions:* Comments on threads about productivity, mindfulness..."
  ],
  "current_context_of_interaction": [
    "Engaging in a casual conversation about managing routines and productivity"
  ]
}

Resources

data/characters/talkDep.json: 12 validated depression personas, covering range from Minimal to Severe.

Which is particularly useful for:

• Screening Agent Evaluation: Testing depression detection models
• Training Data Generation: Creating synthetic conversations for ML training
• Clinician Training: Practicing clinical interview techniques
• Symptom Variability Studies: Exploring how different symptom presentations manifest in conversation