Generative artificial intelligence (AI) technologies, represented by large language models (LLMs), have gradually permeated various clinical scenarios across medical specialties, including diagnostic and treatment decision-making, research support, and medical education; however, significant challenges remain in terms of optimizing specialty-specific adaptation and validating clinical implementation. Therefore, this study systematically reviews the applications of LLMs in medical specialties, focusing on summarizing performance evidence across various specialty-specific generation tasks, and identifies current challenges and future research directions.

By searching the three core comprehensive databases-PubMed, Scopus, and Web of Science-and manually searching the two world-leading journals, Nature and The Lancet, with a search period limited to 2023 through April 2026, we identified publicly available studies on LLMs in medical specialties. Based on data from the included studies, this study employed a descriptive analysis method to conduct a comprehensive systematic review of LLMs in medical specialties.

Searches of the three core comprehensive databases identified 25 studies meeting the inclusion criteria. Additionally, a manual search of top-tier journals in the field, including Nature, The Lancet, and their subsidiary journals, yielded 11 additional studies meeting the criteria, resulting in a total of 36 included studies. The included studies focused on medical specialty LLMs, covering eight specialties including internal medicine, surgery, ophthalmology, oncology, and traditional Chinese medicine general practice. Monomodal models constituted the majority (75%), while multimodal models accounted for only 25%. Applications were primarily concentrated in clinical diagnosis, treatment assistance, and medical education; however, common issues included insufficient specialty adaptation, inadequate utilization of multimodal data, and a lack of evidence for clinical translation.

LLMs have demonstrated significant application potential in medical specialties. However, issues such as insufficient specialty adaptation, inadequate utilization of multimodal data, and a lack of clinical translation evidence remain prevalent. Future efforts should focus on addressing these core issues by deeply developing specialty-specific LLMs, establishing a unified validation and evaluation framework that is clinically driven, clinician-led, and multimodal, and creating a three-pronged regulatory system that integrates hallucination assessment, explainability, and ethical oversight.