Research On Clinical Text Knowledge Extraction And Application Based On Deep Neural Network

With the development of artificial intelligence technology,the field of medical health has received increasing attention.And,medical digitalization has become a key step in the introduction of emerging artificial intelligence technology in the traditional medical and health field.Before the application of electronic medical record technology,the patient's health information is stored in the form of a paper medical record in the hospital,which is difficult to manage,use and store.The rich medical knowledge in the medical record is also difficult to be effectively retrieved and utilized.The electronic medical record records the patient's medical information and health data in detail,which is a rich medical information resource.Therefore,the use of information extraction related technologies to effectively mine and analyze the use of medical knowledge in electronic medical records is of great significance to the development of medical health.This paper is mainly for the semi-structured clinical text data medical knowledge extraction and application research in electronic medical records.Named entity recognition is a primary and critical step in knowledge extraction.By constructing text features of character level,word level and word boundary level as input,this paper adopts Bi-LSTM—CRF model with better performance in entity recognition task,and the F1 value of medical entity recognition reaches 91.6%.The complex entity relationship between medical entities is another manifestation of medical knowledge.The correct rate of entity relationship judgment directly affects the reliability of medical knowledge.In terms of relationship extraction,aiming at the flexible and complex characteristics of medical texts,this paper designed a CNN model based on self-attention mechanism,and integrated multi-entity features,which achieved good results in medical relationship classification tasks,and F1 values were achieved 83.4%.Finally,this paper uses the co-occurrence relationship between symptom entities and disease entities as a priori medical knowledge to construct a disease-symptom cooccurrence knowledge network,and introduces attention mechanism on this network to learn the disease's representation of symptoms and design.Based on the feedforward neural network-based disease diagnosis model,the accuracy rate of TOP1 reached 57.1% and the accuracy rate of TOP3 reached 91.2%.The shortcomings of this experiment are analyzed,which points out a new direction for future research work.