A Multilayer Network and Random Walk-Based Model for Assisting Pattern Identification in Traditional Chinese Medicine Classics

TIAN Siwei; ZENG Ziling; ZONG Wenjing; NIU Qikai; ZHANG Fengxia; JIA Zihan; GUO Zhuang; WANG Jingai; TONG Lin; ZHANG Huamin; LI Bing

Abstract：Objective To construct a model of Traditional Chinese Medicine Classics assisted differentiation based based on multilayer network and random walk algorithm,and to improve the scientificity and reliability of Traditional Chinese Medicine Classics assisted clinical differentiation.Methods Based on the similarity discrimination of ancient and modern clinical identification knowledge,we screened ancient texts on Type 2 Diabetes Mellitus,and adopted a large language model for named entity recognition.With the fine-tuning of BERT training,we constructed Syndrome element determination model named TCM-zs-Bert.Finally,the multi-layer network and random walk algorithm are integrated to form a classical literature assisted differentiation model-CLASD to distinguish syndromes.Results Conduct named entity recognition on highly relevant ancient texts using the Qwen2.5-1.5B model,five types of entities in ancient texts were identified,including symptoms,tongue images,pulse patterns,etiology and pathogenesis,and treatment principles and methods.The accuracy rate was 0.866,the recall rate was 0.866,and the F1 score was 0.856.When the F1 score and recall rate are greater than or equal to 0.8,the TCM-zs-Bert model,which can input ancient entities and output evidence and probability is used for syndrome element prediction.Such as,input the symptom entity"edible,red yellow urine",the predicted pathogenic factor is fire[heat](0.965 6),and the disease location is stomach(0.988 3).When the entity information of"symptoms+etiology and pathogenesis+therapeutic principles and methods":"not thirsty,frequent urination,increased appetite,hot hands and feet,pain in the back and shoulders,turbid urine,polyuria,increased food intake,frequent urination,emaciation;pathogenic heat scorching the five viscera;nourishing kidney water,moistening body fluids",the predicted pathogenic factor are fire[heat](0.990 3)and Jin(liquid)loss(0.044),and the pathological location is the kidney(0.305 3).In addition,in the multi-layer network,there are 18 nodes in the clinical manifestation layer,57 nodes in the Syndrome element layer and 58 nodes in the syndrome types layer.The connecting edge weights between the layers of the network are the probability of the syndrome element,and the value of the relationship between the syndrome element and syndrome type,respectively.As for the CLASD model discriminant results,when LCBX-1 is inputted,it can infer syndromes such as Exuberant Heat Injuring Body Fluids(0.130 0),Exuberant Heat in the Lung and Stomach(0.086 0),and Deficiency of Both Heart and Lung(0.076 0),among others.LCBX-2 corresponds to Yin deficiency and Yang hyperactivity syndrome(0.565 0),Kidney Yin Yang deficiency syndrome(0.058 0),Liver Kidney deficiency syndrome(0.052 0),etc.Comparing the results obtained from CLASD with modern diagnostic and treatment guidelines,the AUC value was 0.77,indicating that the CLASD model has better performance in classical literature assisted syndrome differentiation research.Conclusion Through machine learning and multi-layer network random walks,syndrome element prediction and syndrome discrimination have been achieved,and a classic assisted syndrome differentiation model has been constructed with the path of"clinical manifestations-classic texts-syndrome element discrimination-syndrome inference".In this study,CLASD model is a substitute for the science and objectivity of classical books to assist in differential diagnosis,and a new method was provided to assist clinical decision-making and realise the translational application of the ancient text knowledge.