International Session 50　　Rhinitis and AI

Large language models(LLMs)are rapidly improving and are increasingly explored for medical applications such as clinical note drafting, discharge summaries, differential diagnosis suggestions, and literature/guideline summarization.However, entering patient information into commercial cloud-based models is generally not acceptable, and such models may not be optimized for Japanese clinical guidelines.Locally deployed LLMs that keep data within an institution are therefore desirable.We compared several commercial LLMs with locally deployable LLMs using questions derived from the Japanese otolaryngology board certification examination.We also evaluated whether retrieval-augmented generation (RAG) , which grounds responses in curated references, improves answer accuracy and reduces hallucinations.Outcomes included correctness, clinical appropriateness, and consistency with Japanese guidelines.Local deployment offers advantages in privacy and governance, while RAG may improve reliability by providing evidence-based grounding, especially for local models.Secure, guideline-aligned LLM workflows could enable safer adoption of LLMs in otolaryngology without compromising patient confidentiality.