COIOSDeep risk insights through multi-disciplinary research across the globe
street lights at night
Item new Preprint · not peer-reviewed Added

Differential diagnosis by an LLM agent matched paediatric emergency physicians

An LLM agent and four paediatric emergency physicians gave similarly accurate differential diagnoses for 100 children later admitted.

Dr CaBot, a large language model agent, and four attending paediatric emergency physicians each listed up to five ranked diagnoses. Records came from 100 children seen in one urban paediatric emergency department and then admitted; median age was 4 years. Two reviewers masked to source scored each list 0 to 5 on the Bond scale (weighted kappa 0.64). At first physician evaluation, both scored 4 or more in 95% of cases (difference 0%, 95% CI -5.8 to 5.8%). Performance was also similar at triage, after the full emergency visit and at admission. The study was retrospective, used written records from one specialist centre and did not measure effects on care.

Why it is interesting: Most tests of language models in diagnosis have used adult emergency cases. This one compares an agent with specialist paediatric physicians on real records, with reviewers masked to source.

Source
medRxiv, 18 November 2026
DOI
10.64898/2026.11.16.26367440
Type
Preprint
Design
Retrospective masked comparison, 100 randomly selected admitted paediatric ED encounters at one academic centre, LLM agent (two configurations) vs four attending physicians across four touchpoints, Bond-scored by two blinded reviewers
Verdict
New finding
Driver
AI in medicine

Other drivers in Data, models and systems