| Hospitals have accumulated a large amount of electronic medical record data in the process of medical informatization development,which is a big data source for further comprehensive utilization.Most electronic medical records are unstructured texts,which cannot be directly recognized and analyzed by computers,which seriously affects the processing efficiency.How to parse the effective data information in electronic medical records has become a hot research issue.This paper mainly studies the following two aspects:（1）Named entity recognition is the basic step of text processing.The pre-training models used in the current research are all trained on general texts in the open field.For electronic medical records,which are highly specialized domain texts,they lack background knowledge,resulting in the effect of entity recognition.poor.Aiming at the problem that the pre-trained model lacks medical domain knowledge,this paper designs a named entity recognition algorithm for clinical electronic medical records based on KBERT combined with medical knowledge graph triple dataset.When the electronic medical record is input,the model converts the original clinical electronic medical record sentence into a medical knowledge tree by combining the triple data information of the knowledge graph.At the same time,the visible matrix is used to limit the visible area of each character in the training process,so as to solve the problem of mutual interference between irrelevant words.In this study,the mainstream BIO three-point annotation method was adopted,and three Chinese electronic medical record corpora of main complaint,auxiliary examination and physical examination were constructed and completed in the complex original electronic medical record text.The feasibility of the model is verified on the basis of this real clinical electronic medical record corpus.The experimental results show that the method based on the K-BERT model combined with the medical knowledge graph triple dataset used in this paper achieves 97.99%,97.31%and 96.99% accuracy in the three corpora of main complaint,auxiliary examination and physical examination,respectively.Compared with the algorithm without knowledge graph,it is 0.66%,1.02% and 1.07% higher,respectively.It shows that the algorithm used in this paper is feasible on real data sets.At the same time,this paper also compares with the classical traditional algorithms Bi LSTM+CRF and Bert+Bi LSTM+CRF,and the accuracy rate is also higher than these two algorithms.It also proves that the method based on K-BERT combined with medical knowledge graph has a good effect in the field of clinical electronic medical record named entity recognition.（2）This paper constructs a clinical electronic medical record analysis and management system.According to the application requirements of electronic medical record data analysis and visual display,the system designed the functions of original electronic medical record named entity recognition,user management,disease-related factor analysis and entity labeling.The system also provides filtering functions and user database management functions.In this paper,the useful data in the electronic medical record is screened,and an electronic medical record analysis and management system for doctors is designed. |
