Study On TCM Prescription Recommendation Method Integrating Domain Knowledge

TCM intelligent prescription recommendation refers to learn from medical records of patients and predict candidate Chinese herbs by combining artificial intelligent technologies,so as to simulate the diagnose process of doctors.In recent years,many scholars have carried out related research,but there are still some problems need to be solved in the field of prescription recommendation,such as existing clinical data have the characteristics of "One more and one less",the problem of representing "Unrecorded symptom phenotypes",the performance of existing recommendation methods is relatively low,the rationality of collocation among herbs recommended by existing methods need to be improved,etc.In order to solve the above problems,the following three aspects of TCM prescription recommendation were studied in this work.Firstly,in view of the phenomenon of "One more and one less" existing in clinical case data,a clinical data augmentation method called SOCO was proposed which based on symptom ontology library and symptom co-occurrence relationship,and another data augmentation method called Sab KG was proposed based on knowledge graph sampling.First of all,symptom synonymous relation set,symptom co-occurrence set and herbsymptom knowledge graph were constructed by using symptom ontology library,clinical case data and domain knowledge.After that,two clinical data augmentation methods were proposed.The strategies proposed in this work and baseline methods were applied in the multi-label prescription recommendation task,and the experimental results showed that the two proposed methods can improve the performance of clinical case data,and the Sab KG model achieves relatively optimal performance.Secondly,in view of the performance of existing TCM prescription recommendation methods is relatively low,and the problem of how to represent "Unrecorded symptom phenotypes" better,TCMPR model was proposed for prescription recommendation which based on symptom term mapping and deep learning.In this model,the original symptoms of patients were mapped by subnetwork extraction,and the features of the mapped term set were fused with the constructed symptom network,and the prediction probability of candidate herbs was obtained by learning the constructed neural network model.The performance of TCMPR is compared with that of the baseline methods,and the experimental results showed that the proposed TCMPR achieves the optimal performance,and the method could better represent the "Unrecorded symptom phenotypes" in clinical data by combining domain knowledge.Experiments on key modules of the model showed that the selection of subnetwork screening strategy,symptom representation method,feature fusion method and other modules have impact on the performance of TCMPR model.Thirdly,in view of the rationality of collocation among herbs recommended by existing methods need to be improved,a prescription recommendation method was proposed which based on phenotypic similarity and classical prescription data.In the first place,a prescription recommendation strategy based on classical prescription data was proposed,and experimental results showed the effectiveness of this framework.Considering that this strategy was limited by classical prescription data,another prescription recommendation method based on phenotypic similarity and classical prescription data was proposed.This strategy focused on building phenotypic characteristics of patients,combined domain knowledge such as classical prescription data and symptom network,and formed recommended result by measuring phenotypic similarity between clinical case data and classical prescription data.Experiments were conducted on the proposed patient phenotypic feature construction methods and similarity matching strategies,the results showed that both the SSTM-Sym Jac and SSTM-Sym Cos strategy could effectively improve the recommendation performance,and the compatibility of the recommended results is simultaneously guaranteed.