Loading...
In 2024, I wrote that large language models (LLMs) like GPT-4 could reason through clinical vignettes like experts (NEJM JW Gen Med Jan 15 2025 and NEJM AI 2024; 1:AIcs2400639). Now, researchers from Google have conducted two landmark studies that define the abilities of LLMs in more-complex clinical reasoning.
In one report, the investigators presented >300 New England Journal of Medicine clinicopathologic conference cases (NEJM CPCs) to their LLM model called “Articulate Medical Intelligence Explorer” (AMIE) and to board-certified clinicians. AMIE had access to history of present illness (HPI), and clinicians had access to HPI and test results. AMIE outperformed unassisted physicians in listing the correct diagnosis in the differential dia…