Skip to main content

Large language models are unstable as autonomous medical evaluators across clinical tasks

In brief

Large language model judges under-triage 89% of emergency resuscitation evaluations

Across 10 models, emergency level 1 cases were under-triaged in 89.1% of evaluations, with every model missing these highest-acuity cases at least 70% of the time. Model rankings also shifted across medical tasks, so success on one evaluation task did not predict performance on another. Human oversight and task-specific validation remain essential before clinical use.

Journal
Journal of the American Medical Informatics Association : JAMIA (Q1)
Published
6 November 2026
Study design
Unclassified
Evidence level
Level 5, Expert Opinion (CEBM 5)
Authors
Marcelo Carvalho E Silva, Fernando Carvalho E Silva, Renata de Paula Gomes Nunes de Castro, Alex Gonçalves, Artur Figueiredo Delgado
PMID
42995166
DOI
10.1093/jamia/ocag190

Why clinicians should know about it

  • Picked for Health Informatics (paper of the day, 11 November 2026): Evaluates LLM safety and task-specific validation for clinical triage

Abstract

OBJECTIVE: Large language models (LLMs) are increasingly used as automated evaluators of medical text ("LLM-as-judge"). We tested whether LLM-evaluator performance validated on one medical task against a human reference standard remains stable on a structurally different task. MATERIALS AND METHODS: We evaluated 10 LLMs (3 locally served open-source and 7 API models: Claude, GPT, DeepSeek) as autonomous judges on 3 tasks with human reference labels: LiveQA-Med answer quality (n = 300), PubMedQA research appraisal (n = 300), and Emergency Severity Index (ESI) triage on MIMIC-IV-Ext (n = 256). Each case was scored 3 times per model at temperature zero (25 680 judgments). Analyses used cluster-bootstrap confidence intervals, quadratic-weighted Cohen's κ, and false-discovery-rate correction. RESULTS: Evaluator performance was unstable across tasks: no model ranked in the top 3 on all 3 tasks, and every model shifted by 3 or more rank positions between at least one pair of tasks. In the highest-stakes task, ESI level 1 (Resuscitation) cases were under-triaged in 89.1% of evaluations, in all 10 models (range 70.2%-96.9%). Best triage agreement only reached the human inter-rater range (Claude Opus quadratic-weighted κ = 0.60 [0.54, 0.66]). Locally served 7-27B models (no per-query charge) occupied the Pareto frontier on the text tasks but only its low-concordance end on triage. DISCUSSION: LLM-evaluator performance is unstable between medical tasks that differ in cognitive structure; performance validated on one task should not be assumed to transfer to another. CONCLUSION: Task-specific validation against human reference labels is required before deployment, and emergency triage demands human oversight at the highest-acuity tier.

Abstract as published, via PubMed.

View on PubMedFull text at the publisherOpen in the app

For healthcare professionals. The summary is generated by AI from the published abstract, and the evidence level is assigned automatically from the study design on the Oxford CEBM hierarchy. Neither is medical advice. Read the full paper before changing practice.