Evaluating Large Language Models for Burning Mouth Syndrome Diagnosis

Large language models have been proposed as diagnostic aids across various medical fields, including dentistry. Burning mouth syndrome, characterized by burning sensations in the oral cavity without identifiable cause, poses diagnostic challenges. This study explores the diagnostic accuracy of large language models in identifying burning mouth syndrome, hypothesizing potential limitations.

Clinical vignettes of 100 synthesized burning mouth syndrome cases were evaluated using three large language models (ChatGPT-4o, Gemini Advanced 1.5 Pro, and Claude 3.5 Sonnet). Each vignette included patient demographics, symptoms, and medical history. Large language models were prompted to provide a primary diagnosis, differential diagnoses, and their reasoning. Accuracy was determined by comparing their responses with expert evaluations.

ChatGPT and Claude achieved an accuracy rate of 99%, while Gemini's accuracy was 89% (p < 0.001). Misdiagnoses included Persistent Idiopathic Facial Pain and combined diagnoses with inappropriate conditions. Differences were also observed in reasoning patterns and additional data requests across the large language models.

Despite high overall accuracy, the models exhibited variations in reasoning approaches and occasional errors, underscoring the importance of clinician oversight. Limitations include the synthesized nature of vignettes, potential over-reliance on exclusionary criteria, and challenges in differentiating overlapping disorders.

Large language models demonstrate strong potential as supplementary diagnostic tools for burning mouth syndrome, especially in settings lacking specialist expertise. However, their reliability depends on thorough patient assessment and expert verification. Integrating large language models into routine diagnostics could enhance early detection and management, ultimately improving clinical decision-making for dentists and specialists alike.