Skip to main content

PatientZero

Scaling Synthetic Patient Agents to Real-World Distributions without Real Patient Data

Paper: Arxiv

Overview​

PatientZero role-plays a patient from a clinically grounded synthetic record generated without real patient data. Each record follows the paper's static definition P = {B, S, E} — background patient profile, symptom trajectory, and examination results — sampled from disease knowledge and attribute priors across controlled demographic and severity distributions.

The character file consumed by this client is produced by the PatientZero generator, which standardizes disease knowledge, samples attributes, and validates each case before saving. See that page for the generation pipeline, supported diseases, and how to add a new disease.

Key Features​

  • Disease-grounded: Records are built from standardized disease outlines and disease-specific attribute priors.
  • Distribution control: Demographics, severity, and lifestyle are sampled from global and disease-specific priors.
  • Rich clinical detail: Each case carries symptoms, a mental status exam, scale assessments, and a risk assessment.

Usage​

CLI​

patienthub simulate client=patientZero

Simulate a specific record in the file by index:

patienthub simulate client=patientZero client.data_idx=0

Python​

from patienthub.clients import get_client

client = get_client(agent_name='patientZero', lang='en')
response = client.generate_response("What brings you in today?")
print(response.content)

Configuration​

OptionTypeDefaultDescription
prompt_pathstrdata/prompts/client/patientZero.yamlPath to the role-play prompt file
data_pathstrdata/characters/patientZero.jsonPath to the generated character file
data_idxint0Index of the record in the file

Common API-model options (model_type, model_name, temperature, max_tokens, max_retries, lang) are inherited from the shared client configuration — see the overview.

Character Data Format​

data/characters/patientZero.json is a JSON array of synthetic patient records. The role-play client uses patient_profile, symptom_trajectory, and examination_results; generator metadata may also be present. The full schema and an annotated example are documented in the PatientZero generator → Output Format.

To produce or extend this file, use the generator:

from omegaconf import OmegaConf
from patienthub.generators import get_generator

config = OmegaConf.create({
    "agent_name": "patientZero",
    "disease_key": "depression",
    "output_path": "data/characters/patientZero.json",
})
generator = get_generator(agent_name="patientZero", configs=config, lang="en")
generator.generate_character()

See Also​