Trustworthy large language models for clinical diagnosis: a medical informatics review of evidence, safety, validation, and deployment readiness.

Background

Large language models (LLMs) and multimodal foundation models are increasingly proposed for diagnostic decision support, symptom triage, report interpretation, and electronic health record (EHR)-grounded workflows. Because medical informatics is concerned with how information and communication technologies behave inside clinical practice, workflow evaluation and deployment readiness are central rather than optional.

Objective

To synthesise diagnostic LLM evidence through a medical-informatics lens and to propose a stage-gated framework, with an accompanying self-assessment instrument, for trustworthy clinical use.

Methods

We conducted a critical narrative review supported by targeted evidence mapping of 64 cited records published or updated primarily between 2019 and August 2026, with selected foundational sources retained when directly relevant and emphasis on LLM evidence from 2022 onward. No pooled estimates were calculated.

Results

Twenty empirical diagnostic or patient-facing evaluations were mapped to an eight-level evidence ladder: two at Level 1 (benchmark or examination), one spanning Levels 1-2, ten at Level 2 (case vignettes or synthetic reasoning), three at Level 3 (retrospective clinical records), and four at Level 6 (clinician-in-the-loop workflow), one of which was a prospective single-centre deployment that also reported operational outcomes. Published examples of bias measurement, improvement strategies, and implementation-oriented outcomes exist, but external validation, prospective silent-mode evaluation, controlled clinical-impact trials, and post-deployment surveillance remain rare. Recurring risks include hallucination, dangerous omission, premature closure, unsafe reassurance, poor calibration, automation bias, privacy leakage, prompt injection, and model drift. ConclusionsDiagnostic LLMs should be evaluated as clinical information-system components. We propose a five-gate deployment-readiness framework requiring intended-use specification, evidence matched to clinical risk, a tested safety case, workflow and interoperability readiness, and post-market surveillance, operationalised as a 29-item, use-case-stratified self-assessment instrument.