Abstract
The rapid integration of large language models (LLMs) into healthcare offers significant potential for improving diagnosis, treatment planning, and patient engagement. However, it also presents serious ethical challenges that remain incompletely addressed. In this review, we analyzed 27 peer-reviewed studies published between 2017 and 2025 across four major open-access databases using strict eligibility criteria, robust synthesis methods, and established guidelines to explicitly examine the ethical aspects of deploying LLMs in clinical settings. We explore four key aspects, including the main ethical issues arising from the use of LLMs in healthcare, the prevalent model architectures employed in ethical analyses, the healthcare application domains that are most frequently scrutinized, and the publication and bibliographic patterns characterizing this literature. Our synthesis reveals that bias and fairness ( n = 7 , 25.9%) are the most frequently discussed concerns, followed by safety, reliability, transparency, accountability, and privacy, and that the GPT family predominates ( n = 14 , 51.8%) among examined models. While privacy protection and bias mitigation received notable attention in the literature, no existing review has systematically addressed the comprehensive ethical issues surrounding LLMs. Most previous studies focus narrowly on specific clinical subdomains and lack a comprehensive methodology. As a systematic mapping of open-access literature, this synthesis identifies dominant ethical patterns, but it is not exhaustive of all ethical work on LLMs in healthcare. We also synthesize identified challenges, outline future research directions and include a provisional ethical integration framework to guide clinicians, developers, and policymakers in the responsible integration of LLMs into clinical workflows.