Construction And Application Of Medical Knowledge Graph

Whether in daily life or in the field of professional medical research,the topics and researches related to rational drug use(RDU)have attracted more and more attention.In the era of big data and artificial intelligence(AI),RDU has also become a hot-spot in the cross field of intelligent medical treatment.The implementation of intelligent RDU application is inseparable from a large-scale medical knowledge graph(MKG).However,the existing MKG has single concept or relation and coarse knowledge granularity,which cannot meet the demands of RDU.Furthermore,the construction of a MKG for RDU requires processing a large number of medical texts.Because of the various domain concepts,uneven information and professional expressions in medical texts,the existing knowledge extraction methods are difficult to complete this work.To solve these problems,this dissertation studies the construction technologies of MKG and implements the application of RDU.The main work includes the following aspects:1.An ontology of MKG for RDU is developed,which make a structured modeling for medical knowledge.This ontology includes 41 concepts such as drugs and doses,2 general relations and 31 domain relations.Based on this ontology,the specification of data annotation for drug instruction text is designed.After the annotation of named entities and relations,a medical dataset with 28 categories of entity and 11 categories of relation,a total of 22,600 entities and 7,095 relations,is built.2.A named entity recognition(NER)model,CMG-CRF,for Chinese medical text is proposed,which integrates context and multi-granularity information.It utilizes Bi GRU and Stack CNN to capture the contextual features and features with different granularities of text respectively,and then uses CRF to recognize medical named entities.By experiments on the dataset built in this dissertation,the results show that the F1-score of the proposed model is improved by 1.13% on average comparing with various baseline models.3.Based on the recognition of medical NER,a multi-feature fusion relation extraction(RE)model,MF-PCNN,for Chinese medical text is proposed.It utilizes CRP-EMB to integrate character features,captures the features of entity location and the local features of text by PCNN,and then classifies the text by combining the entity categories.Compared with other baseline models,the proposed model has the best performance on the dataset built in this dissertation.4.Based on the results of NER and RE,and combined with the ontology,a MKG is constructed after the normalization,de-duplication and alignment of entities,which contains356,347 entities and 3,961,541 triples.On this basis,the RDU assistant system is designed and implemented,and the experimental results of the samples verify the rationality and effectiveness of the constructed MKG.