Towards autonomous medical artificial intelligence agents

https://www.nature.com/articles/s41586-026-10675-5

Publish Date: 2026-06-17 18:55:00

Here is a summary of the key points from the article about the MIMIC-IV dataset:

• The dataset derives from MIMIC-IV, a large database of electronic health records from patients at Beth Israel Deaconess Medical Center. It includes clinical information like diagnoses, lab results, medications, procedures, and imaging.

• The study focuses on creating a benchmark dataset of 574 patients with 8 target diagnoses, including abdominal conditions like appendicitis and pancreatic cancer as well as respiratory and urinary infections.

• The dataset curation process involves restricting to admissions for the target diagnoses, removing cases with incomplete data, and randomly selecting 574 cases for physician review and final inclusion in the benchmark.

• A multi-agent system consisting of a patient agent and MIRA (medical AI physician agent) is developed for data retrieval and decision making. MIRA can call tools to request lab tests, imaging, medication prescriptions, etc.

• Evaluations are performed comparing MIRA’s performance to reference data from MIMIC-IV as well as human physicians. Metrics include diagnostic accuracy, procedure evaluation, guideline adherence, and safety.

• Results show MIRA performs comparably to human physicians in diagnosis, procedure identification, medication reconciliation, and guideline following, demonstrating the potential of AI systems in supporting clinical decision making.

That covers the key aspects of the study in a condensed form. Let me know if you need any clarification or have additional questions!