A Study on Similarity Discrimination Method for Ancient and Modern Syndrome Differentiation Knowledge of Type 2 Diabetes Mellitus Based on Pre-trained Language Models

TIAN Siwei; NIU Qikai; LIU Boning; ZENG Ziling; JIA Zihan; ZHANG Fengxia; ZHANG Huamin; LI Bing

Abstract：Objective To develop a method for calculating and discriminating the similarity between ancient and modern syndrome differentiation knowledge,enabling the integration of such knowledge and enhancing the clinical application of classical literature.Methods Taking type 2 diabetes mellitus as an example,datasets of ancient and modern syndrome differentiation literature and clinical simulation data were constructed.Pre-trained language models were applied to calculate the similarity of knowledge between ancient and modern syndrome differentiation,and their performance in processing ancient texts was evaluated.The model with the best performance was selected to compute the similarity between clinical simulation data and original ancient texts,and similarity discrimination rules were established to screen ancient texts with high relevance to clinical simulation data.Results The study identified 36 search terms for ancient disease names related to type 2 diabetes,incorporating 3,850 original texts from ancient books,mainly consisting of disease-related sentences or paragraphs,including disease names,symptoms,etiology and pathogenesis,treatment principles and methods,etc.Modern diagnostic and treatment guidelines include 58 syndrome types and 7 concurrent syndromes.Based on the evaluation of 20 clinical simulation datasets,the TCMLLM-PR model with superior performance was selected to calculate the similarity between clinical simulation data and ancient texts.Based on the short text threshold of≥0.7 and the long text threshold of≥0.6,a total of 2,908 ancient texts highly correlated with clinical simulation data were ultimately screened.Conclusion This study developed a similarity calculation and discrimination method based on pre-trained language models,effectively screening ancient texts related to the clinical manifestations of type 2 diabetes mellitus,thereby providing an effective approach for the integration of ancient and modern syndrome differentiation knowledge and clinical digital intelligence applications.