DynamiCare: A Dynamic Multi-Agent Framework for Interactive and Open-Ended Medical Decision-Making

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Computational Medicine

Background:

• Current AI frameworks for medical decision-making often simulate single-turn tasks, unlike real-world diagnostics.
• The iterative and uncertain nature of clinical diagnosis requires dynamic simulation capabilities.

Purpose of the Study:

• To introduce MIMIC-Patient, a dataset for dynamic, patient-level simulations using EHR data.
• To propose DynamiCare, a dynamic multi-agent framework for interactive clinical diagnosis.
• To establish a benchmark for dynamic clinical decision-making using LLM-powered agents.

Main Methods:

• Developed MIMIC-Patient dataset from MIMIC-III EHRs for patient-level simulations.
• Proposed DynamiCare, a dynamic multi-agent framework modeling diagnosis as an iterative loop.
• Utilized specialist AI agents that query, integrate information, and adapt strategies dynamically.

Main Results:

• Demonstrated the feasibility and effectiveness of the DynamiCare framework through extensive experiments.
• Established the first benchmark for dynamic clinical decision-making with LLM-powered agents.
• Showcased the capability of agents to handle uncertainty and adapt in multi-round interactions.

Conclusions:

• DynamiCare offers a novel approach to simulating dynamic clinical diagnosis.
• The framework and dataset advance research in AI-driven medical decision-making.
• This work sets a new standard for evaluating LLM agents in complex healthcare scenarios.