M3 System Enables Natural Language Queries of MIMIC-IV Clinical Database

A team of researchers has introduced M3, a system designed for natural language queries of the Medical Information Mart for Intensive Care (MIMIC-IV), which is among the largest open-source electronic health record databases globally. Historically, accessing MIMIC-IV necessitated expertise in SQL and clinical knowledge. M3 utilizes the Model Context Protocol to facilitate queries in plain English. It can retrieve MIMIC-IV from PhysioNet with a single command, either by launching a local SQLite instance or connecting to hosted BigQuery. In evaluations against the EHRSQL 2024 benchmark, Claude Sonnet 4 achieved an accuracy of 94%, while the open-weights gpt-oss-20B reached 93% on a set of one hundred answerable questions. The system's goal is to enhance secure access to clinical data for researchers.