系统性疾病与牙髓诊断的关联：使用本地部署大语言模型的回顾性病历研究

Objectives: The objectives of this study were as follows: (1) statistical investigation of the association between systemic diseases and dental pulp condition; and (2) utilization of locally running large language models (LLMs) to handle a large number of dental charts written in natural languages, while protecting patients' information. Methods: A total of 5820 endodontic charts, including healthy controls and patients with a history of cancer, HIV, type 1 diabetes mellitus (DM), and type 2 DM, were selected from 15,147 samples. Using a local LLM environment, the dental charts from each group were classified into six categories: normal pulp (NP); reversible pulpitis (RP); asymptomatic irreversible pulpitis (AIP); symptomatic irreversible pulpitis (SIP); pulp necrosis (PN); and "others", including unclear diagnoses. The distributions of these diagnoses across groups were statistically analyzed. Results: The prompt for the local LLM was repeatedly refined by classifying 400 test cases, resulting in 97.0% agreement with the judgments of three human dentists. The refined prompt was then applied to classify 5820 charts, followed by statistical analyses. Using a generalized linear mixed-effects model (GLMM), the cancer-history group showed a significantly higher NP and a lower AIP proportion, whereas the type 2 DM group showed a significantly lower AIP and a higher SIP proportion. Conclusions: The findings in the cancer-history group could be attributable to a worsened oral environment in the patients and difficulties in diagnosing AIP. Meanwhile, the lower proportion of AIP and higher proportion of SIP in type 2 DM group may reflect pulpal inflammation associated with the disease. Additional attention may be warranted for caries' prevention and treatment in patients with type 2 DM.