The diagnosis of vestibular disorders, such as vestibular migraine, Ménière's disease, and benign paroxysmal positional vertigo (BPPV), often relies heavily on nuanced interpretation of patient history. Variability in clinician experience and access to specialists can lead to diagnostic delays. Artificial intelligence, particularly large language models trained on vast textual corpora, may offer a novel method to extract diagnostic signals from unstructured clinical narratives.

This research, explores the application of state-of-the-art LLMs in a zero-shot learning paradigm for vestibular diagnosis. This study presented standardized, de-identified patient history texts, mimicking clinical notes, to several LLMs without any prior disease-specific fine-tuning. The models were tasked with generating a differential diagnosis. The results show that LLMs could achieve diagnostic accuracy comparable to non-specialist clinicians, effectively distinguishing between disorders based solely on descriptive textual features like symptom quality, timing, and triggers.

This proof-of-concept work suggests that LLMs could serve as a scalable, preliminary decision-support tool, helping triage cases or suggesting diagnostic possibilities in primary care or telemedicine settings. It emphasizes the latent diagnostic value embedded in the patient's own "history." However, significant hurdles must be overcome before clinical integration, including rigorous validation in real-world settings, mitigation of model biases, ensuring patient data privacy, and defining a safe, adjunctive role for AI within the clinician-patient relationship.
DOI：10.15302/ENTD.2025.090005