Psyche

A Multi-faceted Patient Simulation Framework for Evaluation of Psychiatric Assessment Conversational Agents

Paper: arXiv

Generates comprehensive MFC (Multi-Faceted Case) psychiatric profiles for use with the Psyche client agent.

Overview

Property	Value
Key	psyche
Type	LLM-based
Output	Psyche MFC character files

Key Features

• MFC-Profile: Structured clinical intake — demographics, chief complaint, present illness, psychiatric/medical history, medication, family history, and risk assessment
• MFC-History: First-person biography-style life narrative grounded in the profile
• MFC-Behavior: Mental Status Examination — appearance, mood, affect, thought process, insight, and more

How It Works

generate_character(seed) runs three sequential LLM calls:

1. MFC-Profile — generates structured clinical data from the seed record (diagnosis, age, sex)
2. MFC-History — generates a first-person life narrative grounded in the profile
3. MFC-Behavior — generates a Mental Status Examination based on profile and history

The combined MFC object is returned; the generate CLI saves it.

Usage

Provide seeds as a JSON list at data/seeds/psyche.json and run the CLI:

patienthub generate generator=psyche input_path=data/seeds/psyche.json

Each character is written to data/characters/psyche.json (override with output_path).

Configuration

Parameter	Type	Default	Description
agent_name	string	psyche	Generator identifier
prompt_path	string	data/prompts/generator/psyche.yaml	Path to prompt file
model_type	string	"OPENAI"	Model provider key
model_name	string	"gpt-4o"	Model identifier
temperature	float	0.7	Sampling temperature
max_tokens	int	8192	Max response tokens
max_retries	int	3	API retry attempts

Seed Record Format

Seeds live in data/seeds/psyche.json as a JSON list. Each record is validated against PsycheSeed before generation:

[
  {
    "diagnosis": "Major Depressive Disorder",
    "age": 40,
    "sex": "Female"
  }
]

Field	Type	Description
diagnosis	string	Target diagnosis
age	int	Patient age
sex	string	Patient sex

Output Format

{
  "MFC-Profile": {
    "Identifying data": {
      "Age": "40",
      "Sex": "Female",
      "Marital status": "Married",
      "Occupation": "Office worker"
    },
    "Chief complaint": {
      "Description": "I feel overwhelmingly sad and have no energy."
    },
    "Present illness": {
      "Symptom": {
        "Name": "Persistent sadness",
        "Length": 24,
        "Stressor": "work"
      }
    },
    "Past psychiatric history": { "Presence": "No", "Description": null },
    "Impulsivity": {
      "Suicidal ideation": "High",
      "Suicidal plan": "Presence",
      "Homicide risk": "Low"
    }
  },
  "MFC-History": "I grew up in a small town...",
  "MFC-Behavior": {
    "Mood": "Depressed",
    "Affect": "Restricted, anxious, slightly tense",
    "Verbal productivity": "Decreased",
    "Thought process": "Normal"
  }
}

Use Cases

• Creating character files for Psyche client agent simulations
• Psychiatric assessment training datasets
• Evaluating clinical interview agents