Towards autonomous medical artificial intelligence agents

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Electronic Health Records

Background:

• Large language models (LLMs) show promise for clinical decision-making but are often limited to specific tasks.
• Integrating LLMs into electronic health records (EHRs) requires governed data access and safety constraints for clinical workflows.
• The performance of AI agents in managing complex patient cases within an EHR environment remains largely unproven.

Purpose of the Study:

• To evaluate the capabilities of an autonomous AI agent, MIRA, operating within a sandboxed EHR environment.
• To determine if MIRA can achieve physician-level performance in managing patient cases, including data retrieval, test ordering, diagnosis, and treatment planning.
• To compare MIRA's performance against human physicians in terms of diagnostic accuracy and clinical decision-making.

Main Methods:

• MIRA, an autonomous AI agent, was developed to operate within a simulated EHR environment.
• The AI navigated a broad clinical action space, including obtaining patient histories, ordering and interpreting tests, generating differential diagnoses, and formulating treatment plans.
• Simulations were conducted using real patient cases across various diagnoses, comparing MIRA's performance to that of physicians.

Main Results:

• MIRA demonstrated superior diagnostic accuracy compared to physicians in simulated patient cases.
• The AI made guideline-concordant, medication-safe, and appropriate admission decisions.
• MIRA successfully translated clinical intent into structured, actionable EHR operations, outperforming previous LLM applications focused on isolated tasks.

Conclusions:

• An EHR-integrated AI agent like MIRA can effectively manage patient cases and outperform physicians in specific metrics.
• MIRA shows potential as a decision-support partner by enabling structured, actionable EHR operations.
• Further prospective, real-world studies are necessary to confirm generalization, safety, and governance of such AI systems.