Study On Interpretable Prescription Recommendation Method Based On Knowledge Graph

Traditional Chinese medicine(TCM)is a traditional medical system with unique advantages and characteristics in China.As a successful case of traditional Chinese medicine,classic medical records are recorded,which can provide theory support for clinical diagnosis and treatment.However,in clinical,due to the diversity of concept expression of different doctors,it is difficult to map new expression to existing concept due to the limitation of concept library constructed in the past.Therefore,the concept mapping of symptoms becomes the key step to obtain the characteristics of patients.From the ancient and modern Chinese medicine books and other multi-channel knowledge relation,the knowledge graph constructed can establish the relationship between symptoms,drugs,etc.Based on the large-scale data of TCM medical records and knowledge graph,combined with representation learning,information recommendation,clustering and other methods,this paper carried out the research on TCM prescription recommendation methods:(1)Based on the relationship between symptoms,TCM characteristics and Chinese Materia Medica,knowledge graph was constructed.There are 116692 relationships and19605 nodes.This paper proposes a meta path based symptom relation screening and constructs a symptom word network.(2)After screening 5748 cases of TCM medical record data,through the data preprocessing of human-computer collaborative phenotype spectrum tagging system,the symptoms and TCM part of the medical record were analyzed.Research on the method of concept mapping of symptom words,including word segmentation method,construction of subnet,etc.The results show that the correlation between symptom set and prescription drugs can still be maintained after the concept mapping of symptom words,and the feature generated based on subnet construction has the best performance.(3)Based on multi classification,deep neural network prescription recommendation method research,convolutional neural network structure has the best performance.The prescription similarity network and prescription network were constructed.The results about prescription category show that the model with convolution neural network structure has the best performance.(4)Explanatory analysis.Including attention mechanism,shortest path generation.