MedAgentSim Prototype

A simplified implementation of the "Self-Evolving Multi-Agent Simulations for Realistic Clinical Interactions" system described in the paper by Almansoori et al. This prototype uses the meta-llama/Llama-3.2-3B-Instruct model from Hugging Face to simulate doctor-patient interactions in a clinical setting.

Overview

This project simulates a multi-agent clinical environment with:

• Patient Agent: Responds to doctor questions based on a predefined condition.
• Doctor Agent: Engages in multi-turn conversations, requests tests, and makes diagnoses.
• Measurement Agent: Provides test results when requested.

The simulation includes basic conversation and diagnosis phases with a simplified memory system, built using the Llama-3.2-3B-Instruct model.

Requirements

• Python 3.8+
• PyTorch
• Transformers library from Hugging Face
• Access to meta-llama/Llama-3.2-3B-Instruct (Hugging Face token may be required)

Installation

1. Install dependencies:

   pip install torch transformers

Usage

Run the simulation:

python medagentsim_prototype.py

The script will:

1. Initialize a patient with a random condition (fever or chest pain).
2. Simulate 3 turns of doctor-patient conversation.
3. Request a medical test and provide results.
4. Output a diagnosis based on the conversation and test results.

Features

• Multi-turn doctor-patient dialogue.
• Test request and result simulation.
• Basic memory storage for conversation history and diagnoses.
• Chain-of-thought reasoning for diagnosis (simplified).

Limitations

• Uses a smaller model (Llama-3.2-3B) compared to the paper's larger models (e.g., LLaMA 3.3 70B).
• Limited to two conditions (fever, chest pain) in the knowledge base.
• No visual agent support or full self-improvement mechanisms.
• Simplified memory system without ensembling or reflection phases.

Extending the Project

• Add more conditions to MEDICAL_KNOWLEDGE.
• Implement full Experience Replay with reflection and ensembling.
• Integrate a vision model (e.g., LLaVA) for image-based tests.
• Increase conversation turns or enhance prompt engineering.

License

This is a prototype for educational purposes, inspired by the MedAgentSim paper. No official license is provided.

Acknowledgments

Based on the paper "Self-Evolving Multi-Agent Simulations for Realistic Clinical Interactions" by Mohammad Almansoori, Komal Kumar, and Hisham Cholakkal (Mohamed bin Zayed University of Artificial Intelligence).