intelpreventativehealthcare.py

# The code is a simulation of a healthcare system that uses AI agents to manage patient outreach
# Author: Benjamin Consolvo
# Originally created in 2025
# Original code and idea from Mike Lynch on Medium here. Heavily modified.
# https://medium.com/@micklynch_6905/hospitalgpt-managing-a-patient-population-with-autogen-powered-by-gpt-4-mixtral-8x7b-ef9f54f275f1
# https://github.com/micklynch/hospitalgpt

import os
import asyncio
import pandas as pd
import json
import argparse
from typing import Callable, Dict, Any
from autogen import (
    AssistantAgent,
    UserProxyAgent,
    config_list_from_json,
    GroupChat,
    GroupChatManager,
    register_function,
)
from openai import OpenAI
from prompts.epidemiologist_prompt import EPIDEMIOLOGIST_PROMPT
from prompts.doctor_critic_prompt import DOCTOR_CRITIC_PROMPT
from prompts.user_proxy_prompt import USER_PROXY_PROMPT
from prompts.outreach_email_prompt import OUTREACH_EMAIL_PROMPT_TEMPLATE
import aiofiles  # For asynchronous file writing
import functools  # For wrapping synchronous functions in async

# Export the prompt variables for use in the app
__all__ = [
    "get_configs", "target_patients_outreach", "find_patients", 
    "write_outreach_emails", "USER_PROXY_PROMPT", "EPIDEMIOLOGIST_PROMPT", 
    "DOCTOR_CRITIC_PROMPT", "OUTREACH_EMAIL_PROMPT_TEMPLATE"
]

def get_configs(
    env_or_file: str,
    filter_dict: Dict[str, Any]
    ) -> Dict[str, Any]:
    """
    Load configuration from a JSON file.

    Args:
        env_or_file (str): Path to the JSON file or environment variable name.
        filter_dict (Dict[str, Any]): Dictionary to filter the configuration file.

    Returns:
        Dict[str, Any]: Filtered configuration dictionary.
    """
    return config_list_from_json(env_or_file=env_or_file, filter_dict=filter_dict)

async def target_patients_outreach(
    target_screening: str,
    config_list_llama: Dict[str, Any],
    config_list_deepseek: Dict[str, Any],
    log_fn=None,
    user_proxy_prompt=USER_PROXY_PROMPT,
    epidemiologist_prompt=EPIDEMIOLOGIST_PROMPT,
    doctor_critic_prompt=DOCTOR_CRITIC_PROMPT
    ) -> str:
    """
    Determines the criteria for patient outreach based on a screening task.

    This function facilitates a conversation between a user, an epidemiologist, 
    and a doctor critic to define the criteria for patient outreach. The output 
    criteria from the doctor and epidemiologist include minimum age, maximum age, 
    gender, and a possible previous condition.

    Example:

        criteria = asyncio.run(target_patients_outreach("Type 2 diabetes screening"))

    Args:
        target_screening (str): The type of screening task (e.g., "Type 2 diabetes screening").
        config_list_llama (Dict[str, Any]): Configuration for the Llama model.
        config_list_deepseek (Dict[str, Any]): Configuration for the Deepseek model.
        log_fn (callable, optional): Function for logging messages.
        user_proxy_prompt (str, optional): Custom prompt for the user proxy agent.
        epidemiologist_prompt (str, optional): Custom prompt for the epidemiologist agent.
        doctor_critic_prompt (str, optional): Custom prompt for the doctor critic agent.

    Returns:
        str: The defined criteria for patient outreach.
    """
    llm_config_llama: Dict[str, Any] = {
        "cache_seed": 41,
        "temperature": 0,
        "config_list": config_list_llama,
        "timeout": 120,
    }

    llm_config_deepseek: Dict[str, Any] = {
        "cache_seed": 42,
        "temperature": 0,
        "config_list": config_list_deepseek,
        "timeout": 120,
    }

    user_proxy = UserProxyAgent(
        name="User",
        is_termination_msg=lambda x: (
            x.get("content", "") and x.get("content", "").rstrip().endswith("TERMINATE")
        ),
        human_input_mode="NEVER",
        description=user_proxy_prompt,  # Use custom prompt
        code_execution_config=False,
        max_consecutive_auto_reply=1,
    )

    epidemiologist = AssistantAgent(
        name="Epidemiologist",
        system_message=epidemiologist_prompt,  # Use custom prompt
        llm_config=llm_config_llama,
        code_execution_config=False,
        is_termination_msg=lambda x: (
            x.get("content", "") and x.get("content", "").rstrip().endswith("TERMINATE")
        ),
    )

    critic = AssistantAgent(
        name="DoctorCritic",
        system_message=doctor_critic_prompt,  # Use custom prompt
        llm_config=llm_config_deepseek,
        human_input_mode="NEVER",
        code_execution_config=False,
        is_termination_msg=lambda x: (
            x.get("content", "") and x.get("content", "").rstrip().endswith("TERMINATE")
        ),
    )

    groupchat = GroupChat(
        agents=[user_proxy, epidemiologist, critic],
        messages=[]
    )
    manager = GroupChatManager(groupchat=groupchat, llm_config=llm_config_llama)

    user_proxy.initiate_chat(
        manager,
        message=target_screening,
    )
    if log_fn:
        log_fn("Agent conversation complete.")
    user_proxy.stop_reply_at_receive(manager)
    result = user_proxy.last_message()["content"]
    if log_fn:
        log_fn(f"Criteria result: {result}")
    return result

def get_patients_from_criteria(
    patients_file: str,
    min_age: int,
    max_age: int,
    criteria: str,
    gender: str
    ) -> pd.DataFrame:
    """
    Filters patient data from a CSV file based on specified criteria.

    This function reads patient data from a CSV file and filters it based on 
    age range, gender, and a specific condition.

    Example:

        filtered_patients = get_patients_from_criteria(
            patients_file="data/patients.csv",
            min_age=40,
            max_age=70,
            criteria="Adenomatous Polyps",
            gender="None"
        )

    Args:
        patients_file (str): Path to the CSV file containing patient data.
        min_age (int): Minimum age for filtering.
        max_age (int): Maximum age for filtering.
        criteria (str): Condition to filter patients by.
        gender (str, optional): Gender to filter patients by. Defaults to None.

    Returns:
        pd.DataFrame: A DataFrame containing the filtered patient data.
    """
    required_columns = [
        'patient_id', 'First Name', 'Last Name', 'Email',
        'Patient diagnosis summary', 'age', 'gender', 'condition'
    ]

    # Support both file path (str) and file-like object (e.g., from Streamlit)
    if hasattr(patients_file, "read"):
        # Reset pointer in case it's been read before
        patients_file.seek(0)
        patients_df = pd.read_csv(patients_file)
    else:
        patients_df = pd.read_csv(patients_file)

    for column in required_columns:
        if column not in patients_df.columns:
            raise ValueError(f"Missing required column: {column}")

    # Ensure all text is lowercase for case-insensitive matching
    patients_df['condition'] = patients_df['condition'].str.lower()
    criteria = criteria.lower()

    # Filter by condition matching
    condition_filter = patients_df['condition'].str.contains(criteria, na=False)

    # Filter by age range
    age_filter = (patients_df['age'] >= min_age) & (patients_df['age'] <= max_age)

    # Combine filters with OR logic
    combined_filter = age_filter | condition_filter
    
    if gender in ['M', 'F']:
        gender_filter = patients_df['gender'].str.upper() == gender.upper()
        combined_filter = combined_filter & gender_filter

    return patients_df[combined_filter]

def register_function(
    assistant: AssistantAgent,
    user_proxy: UserProxyAgent,
    func: Callable,
    name: str,
    description: str
    ) -> None:
    """
    This function allows an assistant agent and a user proxy agent to execute 
    a specified function.

    Example:
        register_function(
            assistant=assistant_agent,
            user_proxy=user_proxy_agent,
            func=my_function,
            name="my_function",
            description="This is a test function."
        )

    Args:
        assistant (AssistantAgent): The assistant agent to register the function.
        user_proxy (UserProxyAgent): The user proxy agent to register the function.
        func (Callable): The function to register.
        name (str): The name of the function.
        description (str): A description of the function.
    """

    assistant.register_for_llm(
        name=name,
        description=description
        )(func)

    user_proxy.register_for_execution(
        name=name
        )(func)

    return None

async def find_patients(
    criteria: str,
    config_list_llama: Dict[str, Any],
    log_fn=None,
    patients_file_path=None  # Can be a path or a file-like object
    ) -> pd.DataFrame:
    """
    Finds patients matching specific criteria using agents.
    
    This function uses a user proxy agent and a data analyst agent to filter 
    patient data based on the provided criteria.

    Example:
        patients_df = asyncio.run(find_patients(criteria="Patients aged 40 to 70"))

    Args:
        criteria (str): The criteria for filtering patients.
        config_list_llama (Dict[str, Any]): Configuration for the Llama model.
        log_fn (callable, optional): Function for logging messages.
        patients_file_path: Path to patient data file or file-like object.

    Returns:
        pd.DataFrame: A DataFrame containing the filtered patient data.
    """
    # Set up a temporary file path for the agent to use
    temp_file_path = None
    
    # If we have a file-like object (from Streamlit), save it to a temp file
    if patients_file_path is not None and hasattr(patients_file_path, "read"):
        try:
            # Create data directory if it doesn't exist
            os.makedirs("data", exist_ok=True)
            temp_file_path = os.path.join("data", "temp_patients.csv")
            
            # Reset the file pointer and read with pandas
            patients_file_path.seek(0)
            temp_df = pd.read_csv(patients_file_path)
            
            # Save to the temp location
            temp_df.to_csv(temp_file_path, index=False)
            
            if log_fn:
                log_fn(f"Saved uploaded file to temporary location: {temp_file_path}")
                
            # Update the criteria to include the file path
            criteria = f"The patient data is available at {temp_file_path}. " + criteria
        except Exception as e:
            if log_fn:
                log_fn(f"Error preparing patient file: {str(e)}")
            raise
    elif isinstance(patients_file_path, str):
        # It's a regular file path
        temp_file_path = patients_file_path
        criteria = f"The patient data is available at {temp_file_path}. " + criteria
    
    # Configure the LLM
    llm_config_llama: Dict[str, Any] = {
        "cache_seed": 43,
        "temperature": 0,
        "config_list": config_list_llama,
        "timeout": 120,
        "tools": []
    }

    user_proxy = UserProxyAgent(
        name="user_proxy",
        code_execution_config={"last_n_messages": 2, "work_dir": "data/", "use_docker": False},
        is_termination_msg=lambda x: x.get("content", "") and x.get(
            "content", "").rstrip().endswith("TERMINATE"),
        human_input_mode="NEVER",
        llm_config=llm_config_llama,
        # reflect_on_tool_use=True
    )

    data_analyst = AssistantAgent(
        name="data_analyst",
        code_execution_config={
            "last_n_messages": 2, 
            "work_dir": "data/", 
            "use_docker": False},
        llm_config=llm_config_llama,
        # reflect_on_tool_use=True
    )

    register_function(
        data_analyst, 
        user_proxy, 
        get_patients_from_criteria, 
        "get_patients_from_criteria", 
        "Extract and filter patient information based on criteria."
    )
    # --- Fix: Properly extract arguments from the agent conversation ---
    arguments = None  # Ensure arguments is defined in this scope

    def user_proxy_reply(message: str):
        nonlocal temp_file_path
        try:
            if "arguments:" in message:
                arguments_str = message.split("arguments:")[1].strip().split("\n")[0]
                args = eval(arguments_str)
                
                # Override the file path with our temp file if available
                if temp_file_path:
                    args['patients_file'] = temp_file_path
                    if log_fn:
                        log_fn(f"Using patient data from: {temp_file_path}")
                
                return "Tool call received. \nTERMINATE", args
        except Exception as e:
            if log_fn:
                log_fn(f"Error extracting arguments: {e}")
            return f"Error executing function: {str(e)} \nTERMINATE"
        return "Function call not recognized. \nTERMINATE"

    user_proxy.reply_handler = user_proxy_reply
    if log_fn:
        log_fn(f"Set up reply handler with temp file path: {temp_file_path}")

    groupchat = GroupChat(agents=[user_proxy, data_analyst], messages=[])
    manager = GroupChatManager(groupchat=groupchat, llm_config=llm_config_llama)

    chat_output = user_proxy.initiate_chat(data_analyst, message=f"{criteria}")
    user_proxy.stop_reply_at_receive(manager)
    if log_fn:
        log_fn("Agent conversation for patient filtering complete.")

    # Always extract arguments from chat history after chat
    if chat_output and hasattr(chat_output, "chat_history"):
        chat_history = chat_output.chat_history
        for message in chat_history:
            if "tool_calls" in message:
                tool_calls = message["tool_calls"]
                for tool_call in tool_calls:
                    function = tool_call.get("function", {})
                    try:
                        arguments = json.loads(function.get("arguments", None))
                    except Exception:
                        arguments = None
                    if arguments:
                        break
            if arguments:
                break

    if not arguments:
        if log_fn:
            log_fn("Arguments were not populated during the chat process.")
        raise ValueError("Arguments were not populated during the chat process.")

    # Always use the temp file path for the actual data load if available
    if temp_file_path and arguments:
        arguments['patients_file'] = temp_file_path

    filtered_df = get_patients_from_criteria(
        patients_file=arguments['patients_file'],
        min_age=arguments['min_age'],
        max_age=arguments['max_age'],
        criteria=arguments['criteria'],
        gender=arguments['gender']
    )
    if log_fn:
        log_fn(f"Filtered {len(filtered_df)} patients.")
    return filtered_df, arguments

async def generate_email(openai_client, patient, email_prompt, model):
    """
    Asynchronously generate an email using the OpenAI client.

    Args:
        openai_client (OpenAI): The OpenAI client instance.
        patient (dict): The patient data.
        email_prompt (str): The email prompt to send to the model.
        model (str): The model to use for generation.

    Returns:
        str: The generated email content.
    """
    # Wrap the synchronous `create` method in an async function
    create_completion = functools.partial(
        openai_client.chat.completions.create,
        model=model,  # Use model from the OpenAI client
        messages=[{"role": "user", "content": email_prompt}],
        stream=False,
        seed=42,
        temperature=0  # Ensures a consistent output for email (limiting creativity)
    )
    chat_completion = await asyncio.get_event_loop().run_in_executor(None, create_completion)
    return chat_completion.choices[0].message.content


async def write_email_to_file(file_path, patient, email_content):
    """
    Asynchronously write an email to a file.

    Args:
        file_path (str): The path to the file.
        patient (dict): The patient data.
        email_content (str): The email content to write.

    Returns:
        None
    """
    async with aiofiles.open(file_path, "w") as f:
        await f.write(f"Name: {patient['First Name']} {patient['Last Name']}\n")
        await f.write(f"Patient ID: {patient['patient_id']}\n")
        await f.write(f"Email: {patient['Email']}\n")
        await f.write(email_content)
        await f.write("\n")
        await f.write("-----------------------------------------")


async def write_outreach_emails(
    patient_details: pd.DataFrame,
    user_proposal: str,
    arguments_criteria: Dict[str, Any],
    openai_client: OpenAI,
    model: str,
    phone: str = "123-456-7890",
    email: str = "[email protected]",
    name: str = "Benjamin Consolvo",
    log_fn=None,
    outreach_email_prompt_template=OUTREACH_EMAIL_PROMPT_TEMPLATE
    ) -> None:
    """
    Asynchronously generates and writes outreach emails for patients.

    This function generates personalized emails for patients based on their 
    details and the specified screening criteria. The emails are written to 
    individual text files asynchronously.

    Args:
        patient_details (pd.DataFrame): DataFrame containing patient details.
        user_proposal (str): The type of screening task (e.g., "Colonoscopy screening").
        arguments_criteria (Dict[str, Any]): The criteria used for filtering patients.
        openai_client (OpenAI): The OpenAI client instance.
        model (str): Model name to use for generation.
        phone (str): Phone number to include in the outreach emails.
        email (str): Email address to include in the outreach emails.
        name (str): Name to include in the outreach emails.
        log_fn (callable, optional): Function for logging messages.
        outreach_email_prompt_template (str): Custom template for outreach emails.

    Returns:
        None
    """
    os.makedirs("data", exist_ok=True)
    if patient_details.empty:
        msg = "No patients found"
        print(msg)
        if log_fn:
            log_fn(msg)
        return

    async def process_patient(patient):
        # Ensure all required fields are present in the patient record
        required_fields = ['First Name', 'Last Name', 'patient_id', 'Email']
        for field in required_fields:
            if field not in patient or pd.isna(patient[field]):
                msg = f"Skipping patient record due to missing field: {field}"
                print(msg)
                if log_fn:
                    log_fn(msg)
                return

        # Validate the prompt template
        try:
            # Use the custom template instead of the default
            email_prompt = outreach_email_prompt_template.format(
                patient=patient.to_dict(),
                arguments_criteria=arguments_criteria,
                first_name=patient["First Name"],
                last_name=patient["Last Name"],
                user_proposal=user_proposal,
                name=name,
                phone=phone,
                email=email
            )
        except KeyError as e:
            msg = f"Error formatting email prompt: Missing key {e}. Skipping patient."
            print(msg)
            if log_fn:
                log_fn(msg)
            return

        msg = f'Generating email for {patient["First Name"]} {patient["Last Name"]}'
        print(msg)
        if log_fn:
            log_fn(msg)
        email_content = await generate_email(openai_client, patient, email_prompt, model)

        file_path = f"data/{patient['First Name']}_{patient['Last Name']}_email.txt"
        await write_email_to_file(file_path, patient, email_content)
        if log_fn:
            log_fn(f"Wrote email to {file_path}")

    tasks = [process_patient(patient) for _, patient in patient_details.iterrows()]
    await asyncio.gather(*tasks)

    msg = f"All emails have been written to the 'data/' directory."
    print(msg)
    if log_fn:
        log_fn(msg)

def parse_arguments():
    """
    Parse command-line arguments for the script.

    Returns:
        argparse.Namespace: Parsed arguments.
    """
    parser = argparse.ArgumentParser(description="Run the Preventative Healthcare Intel script.")
    parser.add_argument(
        "--oai_config",
        type=str,
        required=True,
        help="Path to the OAI_CONFIG_LIST.json file."
    )
    parser.add_argument(
        "--target_screening",
        type=str,
        required=True,
        help="The type of screening task (e.g., 'Colonoscopy screening')."
    )
    parser.add_argument(
        "--patients_file",
        type=str,
        default="data/patients.csv",
        help="Path to the CSV file containing patient data. Default is 'data/patients.csv'."
    )
    parser.add_argument(
        "--phone",
        type=str,
        default="123-456-7890",
        help="Phone number to include in the outreach emails. Default is '123-456-7890'."
    )
    parser.add_argument(
        "--email",
        type=str,
        default="[email protected]",
        help="Email address to include in the outreach emails. Default is '[email protected]'."
    )
    parser.add_argument(
        "--name",
        type=str,
        default="Benjamin Consolvo",
        help="Name to include in the outreach emails. Default is 'Benjamin Consolvo'."
    )
    return parser.parse_args()

if __name__ == "__main__":
    # Parse command-line arguments
    args = parse_arguments()

    llama_filter_dict = {"model": ["meta-llama/Llama-3.3-70B-Instruct"]}
    config_list_llama = get_configs(args.oai_config,llama_filter_dict)

    deepseek_filter_dict = {"model": ["deepseek-ai/DeepSeek-R1-Distill-Llama-70B"]}
    config_list_deepseek = get_configs(args.oai_config,deepseek_filter_dict)
    
    # Validate API key before initializing OpenAI client
    api_key = config_list_llama[0].get('api_key')
    
    if not api_key:
        config_list_llama[0]['api_key'] = config_list_deepseek[0]['api_key'] = api_key = os.environ.get("OPENAI_API_KEY")

    # Get the criteria for the target screening
    # The user provides the screening task.
    # The epidemiologist and doctor critic will then define the criteria for the outreach.
    filepath = os.path.join(os.getcwd(), args.patients_file)
    criteria = f"The patient data is located here: {filepath}."
    criteria += asyncio.run(target_patients_outreach(args.target_screening,config_list_llama, config_list_deepseek))
    
    # The user proxy agent and data analyst
    # will filter the patients based on the criteria defined by the epidemiologist and doctor critic.
    patients_df, arguments_criteria = asyncio.run(find_patients(criteria,config_list_llama, patients_file_path=filepath))
    
    # Initialize OpenAI client
    openai_client = OpenAI(
        api_key=api_key,
        base_url=config_list_llama[0]['base_url']
    )
    
    #Use LLM to write the outreach emails to text files.
    asyncio.run(write_outreach_emails(
        patients_df, 
        args.target_screening, 
        arguments_criteria, 
        openai_client,
        config_list_llama[0]['model'],
        phone=args.phone,
        email=args.email,
        name=args.name
    ))