H-AdminSim is an official Python package for simulating interactions between hospital administrative staff and first-visit outpatients using LLM agents. It provides a standardized evaluation testbed for assessing LLM performance across key administrative tasks across multiple care levels (primary, secondary, and tertiary), with optional FHIR integration and support for heterogeneous deployment environments, allowing flexible simulation workflows tailored to diverse hospital systems.
Large hospitals often handle 10,000+ outpatient encounters per day, and prior reports indicate limited specialization among administrative staff despite high workload. H-AdminSim is designed to help address these challenges by offering a realistic, reproducible simulation environment that supports future hospital automation and LLM-assisted administrative workflows.
- Paper: https://arxiv.org/abs/2602.05407
- PyPI Package: https://pypi.org/project/h-adminsim/
We provide configuration examples for simulating primary, secondary, and tertiary care settings. Each configuration reflects key characteristics of its hospital level:
- Hospital time granularity:
tertiary<secondary=primary(coarser in lower levels) - Number of departments:
primary<secondary<tertiary - Number of physicians:
primary<secondary<tertiary - Patient referral rate:
primary<secondary<tertiary - Proportion of patients with preferred
physician/date:primary<secondary=tertiary
You may also define your own conditions using a custom configuration file (e.g., data_synthesis.yaml)
We extend the previously emergency department-focused PatientSim to enable realistic conversations between administrative staff and first-visit outpatients with diverse backgrounds.
- Disease profile: One of 194 disease–symptom pairs across 9 internal-medicine departments (
gastroenterology,cardiology,pulmonology,endocrinology/metabolism,nephrology,hematology/oncology,allergy,infectious diseases,rheumatology) - Medical referral status: Dialogue flow adapts based on whether the patient has a referral
- Tasks: Department recommendation, information extraction, structured data construction
We simulate realistic scheduling interactions between administrative staff and patients, reflecting diverse scheduling behaviors and hospital-level constraints.
- Time flow: Users can define the simulation period and starting point, enabling the agent to perform time-related tasks based on the progression of simulated time.
- Patient preferences:
ASAP(earliest slot),physician(specific physician requested),date(preferred date range start) - Random requests:
cancellation,rescheduling - Tasks: New appointment scheduling, rescheduling, schdule cancellation
We provide optional support for integrating with FHIR, allowing the simulator to operate flexibly across heterogeneous hospital environments as long as FHIR-compatible data is available. For instructions on running a FHIR server, please refer to the FHIR Server Execution repository.
- February 2026 (v1.0.3): Update error exception codes for observable traceback.
- February 2026 (v1.0.2): Update dependencies for stable simulation.
- February 2026 (v1.0.1): H-AdminSim now supports concurrent simulation and open-source tool calling.
- February 2026 (v1.0.0): H-AdminSim package has been released.
- December 2025 (v0.7.2): Rule-based and tool calling-based scheduling logics have been supported.
- November 2025 (v0.7.1): Self-corrective feedback logic has been supported.
- October 2025 (v0.7.0): Simulation has been improved reflecting feedbacks from experts.
- September 2025 (v0.6.0): Simulation has been improved reflecting feedbacks from experts.
pip install h_adminsimimport h_adminsim
print(h_adminsim.__version__)
Before using the LLM API, you need to provide the API key (or the required environment variables for each model) either directly or in a .env file.
# For GPT API without Azure
OPENAI_API_KEY=${YOUR_OPENAI_KEY}
# For Gemini API
GOOGLE_API_KEY=${YOUR_GEMINI_API_KEY"}
from h_adminsim import AdminStaffAgent, SupervisorAgent
from h_adminsim.pipeline import DataGenerator, Simulator
from h_adminsim.task.agent_task import OutpatientFirstIntake, OutpatientFirstScheduling
data_generator = DataGenerator()
data_generator.build(convert_to_fhir=True)
agent_data_dir = data_generator.save_dir / 'agent_data'
output_dir = data_generator.save_dir / 'simulation_results'
# Intake task
intake_task = OutpatientFirstIntake(
patient_model='gpt-5-nano',
admin_staff_model='gemini-2.5-flash',
)
# Scheduling task
scheduling_task = OutpatientFirstScheduling(
patient_model='gpt-5-nano',
admin_staff_model='gpt-5-mini',
)
# Simulation
simulator = Simulator(
intake_task=intake_task,
scheduling_task=scheduling_task,
)
simulator.run(
simulation_data_path=agent_data_dir,
output_dir=output_dir,
resume=False,
verbose=True
)
from h_adminsim.pipeline import DataGenerator
# 1. Generator Initialization
# 1.1. Default usaage
data_generator = DataGenerator() # Default: primary care
# data_generator = DateGenerator(level='secondary') # For secondary care
# data_generator = DateGenerator(level='tertiary') # For tertiary care
# 1.2. You can synthesize data with your own configuration
data_generator = DataGenerator(config='data_config.yaml')
# 2. Synthesizing Data
# 2.1. Default usage
data_generator.build()
# 2.2. When you want the synthesized data returned along with its FHIR-converted version (optional)
data_generator.build(convert_to_fhir=True)
# 2.3. When you want to upload the synthesized data to your own FHIR server (optional)
# Provide your FHIR server URL
data_generator.upload_to_fhir(
fhir_data_dir=data_generator.save_dir / "fhir_data",
fhir_url=${FHIR_URL},
) Configuration example for data synthesis
# Base
seed: 9999
# FHIR server url
fhir_url: http://localhost:8080/fhir # Optional: set your FHIR server URL here
# Data configs
project: ./synthetic_data/
data_name: hospital_small # Output path: ./synthetic_data/hospital_small/data
hospital_data:
hospital_n: 10 # Number of hospitals to synthesize
start_date:
min: 2025-03-17 # ISO format: YYYY-MM-DD
max: 2025-09-21
days: 7 # Simulation period (in days)
interval_hour: 0.25 # Time unit expressed in hours
start_hour: # Possible hospital opening hours
min: 9
max: 10
end_hour: # Possible hospital closing hours
min: 18
max: 19
department_per_hospital:
min: 7
max: 9
doctor_per_department:
min: 1
max: 1
working_days: # Number of days each doctor works during the simulation period
min: 3
max: 4
doctor_capacity_per_hour:
min: 1
max: 4
doctor_has_schedule_prob: 0 # Probability that a doctor has at least one fixed schedule
schedule_coverage_ratio: # Proportion of fixed schedules relative to total working hours
min: 0.4
max: 0.6
appointment_coverage_ratio: # Proportion of appointments scheduled outside fixed schedules
min: 0.2
max: 0.5
preference:
type: ['asap', 'doctor', 'date'] # Types of patient scheduling preferences
probs: [0.4, 0.4, 0.2] # Probability distribution for each preference type
symptom:
type: ['simple', 'with_history'] # 'simple' = no referral; 'with_history' = referral case
probs: [0.7, 0.3] # Probability distribution for symptom types
from h_adminsim import SupervisorAgent
from h_adminsim.task.agent_task import OutpatientFirstIntake
# 1. Patient Intake
# 1.1. Default usage (Staff-only)
intake_task = OutpatientFirstIntake(
patient_model='gpt-5-nano',
admin_staff_model='gpt-5-mini',
intake_max_inference=5, # Default: up to 5 rounds (10 turns) of dialogue
)
##############################################################
# 1.2. Role separation
# Staff: dialogue handling, Supervisor: data collection and structuring
supervisor_agent = SupervisorAgent(
target_task='first_outpatient_scheduling',
model='gemini-2.5-flash',
api_key=${YOUR_API_KEY}, # You may set the API key here instead of using a .env file
)
intake_task = OutpatientFirstIntake(
patient_model='gemini-2.5-flash',
admin_staff_model='gpt-5',
supervisor_agent=supervisor_agent,
intake_max_inference=8,
)
##############################################################
# 1.3. Advanced usage: vLLM
supervisor_agent = SupervisorAgent(
target_task='first_outpatient_scheduling',
model='meta-llama/Llama-3.3-70B-Instruct',
use_vllm=True, # Use a vLLM-hosted model as the supervisor
vllm_endpoint='http://0.0.0.0:8000', # vLLM server endpoint
)
intake_task = OutpatientFirstIntake(
patient_model='meta-llama/Llama-3.3-70B-Instruct',
admin_staff_model='meta-llama/Llama-3.3-70B-Instruct',
supervisor_agent=supervisor_agent,
intake_max_inference=5,
patient_vllm_endpoint='http://0.0.0.0:8000',
admin_staff_vllm_endpoint='http://0.0.0.0:8000',
)
##############################################################
from h_adminsim import AdminStaffAgent, SupervisorAgent
from h_adminsim.task.agent_task import OutpatientFirstScheduling
# 2. Appointment Scheduling
# 2.1. Default usage (Tool-calling with reasoning fallbacks)
scheduling_task = OutpatientFirstScheduling(
patient_model='gpt-5-nano',
admin_staff_model='gemini-2.5-flash',
schedule_cancellation_prob=0.05, # Cancellation event
request_early_schedule_prob=0.1, # Rescheduling event
preference_rejection_prob = 0.3, # Prob. of rejecting the first-priority scheduling preference
preference_rejection_prob_decay = 0.5, # Decay factor for the preference rejection prob.
scheduling_max_inference=5,
scheduling_strategy='tool_calling',
fhir_integration=False,
)
##############################################################
# 2.2. LLM reasoning-based scheduling without tool-calling
scheduling_task = OutpatientFirstScheduling(
patient_model='gpt-5-nano',
admin_staff_model='gpt-5-mini',
schedule_cancellation_prob=0.05, # Cancellation event
request_early_schedule_prob=0.1, # Rescheduling event
preference_rejection_prob = 0.3, # Prob. of rejecting the first-priority scheduling preference
preference_rejection_prob_decay = 0.5, # Decay factor for the preference rejection prob.
scheduling_max_inference=5,
scheduling_strategy='reasoning',
fhir_integration=False,
)
##############################################################
# 2.3. HIS upload via FHIR
scheduling_task = OutpatientFirstScheduling(
patient_model='gpt-5-nano',
admin_staff_model='gemini-2.5-flash',
schedule_cancellation_prob=0.05, # Cancellation event
request_early_schedule_prob=0.1, # Rescheduling event
preference_rejection_prob = 0.3, # Prob. of rejecting the first-priority scheduling preference
preference_rejection_prob_decay = 0.5, # Decay factor for the preference rejection prob.
scheduling_max_inference=5,
scheduling_strategy='tool_calling',
fhir_integration=True,
)
##############################################################
# 2.4. Advanced usage: vLLM
scheduling_task = OutpatientFirstScheduling(
patient_model='meta-llama/Llama-3.3-70B-Instruct',
admin_staff_model='gpt-5-mini',
schedule_cancellation_prob=0.05, # Cancellation event
request_early_schedule_prob=0.1, # Rescheduling event
preference_rejection_prob = 0.3, # Prob. of rejecting the first-priority scheduling preference
preference_rejection_prob_decay = 0.5, # Decay factor for the preference rejection prob.
scheduling_max_inference=5,
scheduling_strategy='tool-calling', # Currently, we do not support tool-calling from vLLM
fhir_integration=False,
patient_vllm_endpoint='http://0.0.0.0:8000',
)
##############################################################
from h_adminsim.pipeline import Simulator
# 3. Simulator initialization
# 3.1. Default usage
simulator = Simulator(
intake_task=intake_task,
scheduling_task=scheduling_task,
simulation_start_day_before=3,
fhir_integration=False,
fhir_url=None,
fhir_max_connection_retries=5,
random_seed=9999,
)
##############################################################
# 3.2. FHIR integration
# (If enabled, scheduling task must be initialized with `fhir_integration=True`)
simulator = Simulator(
intake_task=intake_task,
scheduling_task=scheduling_task,
simulation_start_day_before=3,
fhir_integration=True,
fhir_url='http://localhost:8080/fhir',
fhir_max_connection_retries=5,
random_seed=9999,
)
##############################################################
# 3.3. Running a single task
# 3.3.1. Intake task only
simulator = Simulator(
intake_task=intake_task,
scheduling_task=None,
simulation_start_day_before=3,
fhir_integration=False,
fhir_url=None,
fhir_max_connection_retries=5,
random_seed=9999,
)
# 3.3.2. Scheduling task only
simulator = Simulator(
intake_task=None,
scheduling_task=scheduling_task,
simulation_start_day_before=3,
fhir_integration=False,
fhir_url=None,
fhir_max_connection_retries=5,
random_seed=9999,
)
############################################################### Run the initialized simulator
simulator.run(
simulation_data_path='hospital_data/primary/agent_data',
output_dir='hospital_data/primary/simulation_results',
resume=False, # If the simulation stopped unexpectedly, set resume=True with the same paths
verbose=True,
)
For H-AdminSim and PatientSim outpatient simulation, please cite the following.
@misc{lee2026hadminsimmultiagentsimulatorrealistic,
title={H-AdminSim: A Multi-Agent Simulator for Realistic Hospital Administrative Workflows with FHIR Integration},
author={Jun-Min Lee and Meong Hi Son and Edward Choi},
year={2026},
eprint={2602.05407},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2602.05407},
}