Research On Auxiliary Decision-making Of Traditional Chinese Medicine Syndrome Differentiation And Interpretability

Traditional Chinese Medicine(TCM)has accumulated rich experience over thousands of years of practice,forming a unique medical system and providing abundant data resources for medical research,most of which exist in the form of medical records containing valuable and effective experience of TCM doctors in diagnosing and treating diseases.The profound and long learning cycle of TCM thinking,coupled with a shortage of high-quality TCM resources at the grassroots level and a high demand for TCM services,has led to a contradiction between the increasing demand for TCM specialty diagnosis and treatment services and limited TCM resources.To alleviate this contradiction,deep learning and machine learning technologies can be used to analyze and utilize medical records,assisting TCM practitioners in making syndrome differentiation decisions,effectively improving diagnostic and therapeutic efficiency,developing more reasonable treatment plans,and promoting the sharing and inheritance of TCM clinical knowledge.TCM clinical diagnosis and treatment follow the basic principle of "syndrome differentiation and treatment".Doctors identify the nature and location of the disease through comprehensive analysis of the four diagnostic manifestations of "observation,hearing,inquiry,and cutting",and diagnose a certain syndrome before proceeding with treatment.However,in the real world,the sentence structure of TCM’s four diagnostic texts is diverse,symptoms are expressed differently,and there is a lack of unified clinical terminology standards.There is a general flexibility and personalization in wording and sentence formation,and the dialectical thinking of "four diagnostic integration" in TCM leads to more complex extraction and combination of four diagnostic features.In addition,TCM syndrome lack unified standard specifications.The complex combinations of TCM syndrome and the variability of diagnostic outcomes among individuals have resulted in an excessively sparse labeling space for TCM syndrome,thereby increasing the difficulty of TCM syndrome differentiation.On the other hand,there is a lot of research on TCM syndrome differentiation assistance decision-making,but its clinical application is very limited.One of the key reasons is the insufficient interpretability of syndrome differentiation.Due to the lack of understanding of the principle of model decisionmaking,doctors and patients find it difficult to trust the syndrome differentiation results of the model.In view of the above problems,this paper focuses on researching the interpretability of models for TCM diagnosis and treatment,by utilizing two approaches for TCM diagnostic assistance.Firstly,multi-label classification is employed for direct TCM diagnosis.Secondly,similarity recommendation is used to indirectly diagnose TCM by identifying similar cases.Specific contents include:(1)Regarding the excessively sparse nature of the labeling space for TCM syndrome,a two-stage multi-label classification method for TCM syndrome differentiation was proposed.Firstly,a neural network model was used to classify syndrome elements into multiple labels,and a sparse attention mechanism was employed to capture the relevant keywords and their weights to generate the syndrome element representation.Secondly,a random forest model was used to train the syndrome element representation with relevant label features,and the extracted rules were used to generate the syndrome differentiation patterns,enhancing the interpretability of TCM syndrome differentiation.The experimental results showed that the proposed method improved the accuracy of syndrome differentiation while maintaining a high F1 score,facilitating the exploration of syndrome differentiation patterns and making the decision-making process easier to understand for clinicians.(2)Due to the complexity and diversity of TCM syndromes and diagnosis methods,and the difficulty for doctors to accept the results directly generated by deep models,this study calculates the similarity between TCM four diagnostic texts and recommends existing medical cases with similar symptoms to assist in diagnosing TCM syndromes.To address the flexible and variable sequence of TCM diagnostic texts,a multi-level feature fusion and interpretable method for calculating TCM four diagnostic similarity is proposed.The text is represented from two levels: considering the word sequence and weakening the word sequence.Sparse attention mechanism is used to focus on key features to enhance the interpretability of the model.Gradient boosting tree is then introduced to capture various distinguishing diagnostic features to accurately predict the similarity between the two diagnostic texts.Experimental results show that this method can effectively improve the semantic representation of diagnostic texts,eliminate the influence of irrelevant features,and enhance the capture of the combined features of two diagnostic texts.The recommended medical cases not only have a clear similarity basis but also have TCM syndromes derived from real medical cases,making them more easily understandable and accepted by doctors and increasing the credibility of the diagnosis.(3)Based on the above research results,a Python-based and related development tools are used to design and construct an interpretable TCM syndrome differentiation auxiliary decision-making system,which provides reference for doctors’ syndrome differentiation.