Research On The Application Of Medical Text Based On Deep Learning And Knowledge Graph

Under the general trend of Medical-Engineering integration,the application of artificial intelligence technology to solve related problems in the medical field has become an emerging research direction.This thesis focuses on the practical application scenarios of medical texts and hopes to solve the bottleneck of the industry through natural language processing technology,so as to help the medical field realize informatization,digitalization,and intelligence.This thesis includes the following work:(1)Medical named entity recognition.The models of word segmentation and dictionary,conditional random field,bidirectional long short-term memory networks combined with conditional random field model(Bi-LSTM-CRF)are selected for research.The open source EMR data set is preprocessed,frequency statistics and visual display.Through experimental comparison,it is found that the char-based word2 vec Bi-LSTM-CRF model performs best,with a precision rate of 80% and a recall rate of80%.The model method is used in the following application research of construction medical knowledge graph and medical guidance classification.(2)Construction of medical knowledge graph.The professional medical information website was selected as the data source.After processing through data cleaning,entity recognition and other methods,a total of 22 040 medical knowledge entities and 137 732 entity relationships were obtained.The Neo4 j graph database is used to construct the medical knowledge graph,and some medical entity samples are visualized through query statements.(3)Classification study of medical guidance.A Knowledge Interactive Attention Network(KIAN)medical guidance model is proposed to classify short texts of patients’ chief complaints.The model introduces external medical knowledge entities,and uses attention mechanism to judge the importance and relevance of knowledge entities to short texts.In this thesis,the KIAN-LSTM model of recurrent neural network based on self-attention mechanism and KIAN-CNN model based on convolution neural network are used to test and analyze 24 kinds of patients’ chief complaint text data sets.In addition,five other mainstream text classification methods are selected for comparison.The results show that the comprehensive performance of the knowledge interactive attention network(KIAN)medical guidance model proposed in this thesis is the best,and its classification accuracy is about one to two percentage points higher than other methods.When KIAN-LSTM uses characters or words as input,the accuracy reaches80.65% and 82.65% respectively,and the macro F1 reaches 78.59% and 79.66%respectively.When KIAN-CNN uses characters or words as input,the accuracy reaches84.6% and 85.1% respectively,and the macro F1 reaches 82.45% and 82.71%respectively.