# Large Language Models in Radiologist–Patient Communication: A Narrative Review for Clinical Practice

## TL;DR

This paper reviews how large language models are used in radiology to improve patient communication, but highlights the need for professional oversight and further research.

## Contribution

The paper provides a comprehensive narrative review of LLM applications in radiology, emphasizing safety, effectiveness, and regulatory considerations.

## Key findings

- LLMs improved readability by 2-6 grade levels but required professional review in up to 80% of cases.
- Translation accuracy was higher for high-resource languages, with more errors in low-resource languages.
- Patients accepted AI-assisted communication when clinicians were clearly involved.

## Abstract

Large language models (LLMs) are used in radiology to simplify reports, translate findings, and support patient-facing communication, yet their clinical value and safety remain uncertain. This narrative review was conducted in accordance with the Scale for the Assessment of Narrative Review Articles (SANRA) quality criteria and synthesises evidence from 49 studies published between 2020 and 2025, focusing on clinician-mediated use of LLMs across four domains: report simplification, multilingual translation, patient education, and patient attitudes. Across studies, LLMs consistently improved readability by 2-6 grade levels, but only one randomised trial directly assessed patient comprehension. A professional review was required in up to 80% of outputs in controlled settings, compared with <10% in observational studies. Harmful factual errors were uncommon but non-negligible (0-10% depending on task and model). Translation performance was highest for high-resource languages, while semantic drift was more frequent in low-resource languages, necessitating bilingual review. Patients generally accepted AI-assisted communication when clinician oversight was explicit. Current regulatory and professional guidance support supervised, institution-hosted deployment. Evidence supports specific use cases, patient summaries, translation drafts, and educational materials, but does not justify autonomous deployment or direct patient self-use. Key evidence gaps remain in comprehension outcomes, workflow impact, and real-world validation.