AN ELECTRONIC health record integrated AI agent demonstrates improved diagnostic accuracy and clinical decision making in simulated EHR environments, outperforming physicians across a range of real patient cases. The findings suggest that large language model-based systems may extend beyond isolated clinical tasks towards structured decision support within EHRs.
Diagnostic Accuracy and Clinical Decision Making
The system, identified as MIRA (Medical Intelligence for Reasoning and Action), operates as an autonomous AI agent within a sandboxed EHR environment. It is designed to access patient data under governed conditions and execute permitted clinical actions within defined safety constraints. Within this setting, the EHR-integrated AI agent retrieves patient histories, orders and interprets laboratory investigations, imaging studies and microbiology tests, and generates differential diagnoses alongside structured treatment plans.
In simulated evaluations using real patient cases spanning multiple diagnostic categories, the EHR-integrated AI agent demonstrated higher diagnostic accuracy than physicians. The system also produced guideline concordant recommendations and safe medication decisions, including appropriate admission planning.
The data indicate that the EHR-integrated AI agent was able to navigate a broad clinical action space while maintaining clinically appropriate outputs. Compared with earlier large language model applications that focused on narrow subtasks or free text advice, this system translated clinical intent into structured EHR actions, including ordering investigations and proposing treatment pathways.
Implications for EHR Integrated AI Systems
The findings suggest potential for EHR integrated AI systems to support clinical reasoning and decision making at scale within EHR environments. The EHR integrated AI agent demonstrated the ability to integrate diagnostic reasoning with actionable clinical steps in a simulated setting using real patient data.
The authors expressed that that such systems may be best positioned within collaborative clinical environments, where they support physicians by carrying out routine tasks under defined levels of supervision or by providing evidence based therapeutic recommendations that remain subject to physician approval, thereby augmenting rather than replacing clinical expertise.
However, the evaluation was conducted in simulation within a sandboxed EHR environment. Further research is required to determine performance, safety and reliability in prospective real world clinical settings and across diverse healthcare systems.
Reference
Ferber D et al. Towards autonomous medical artificial intelligence agents. Nature. 2026;DOI:10.1038/s41586-026-10675-5.
Featured Image: sebra on Adobe Stock
- Author:
