Research On Medical Event And Temporal Relation Extraction Technology Based On Chinese Electronic Medical Records

In Chinese electronic medical records,extracting information such as medical events and their temporal relationships to construct a timeline of patient health status is crucial for intelligent medical applications such as disease monitoring,clinical decision-making assistance,and evaluation of drug treatment effects.There have been many studies on the extraction of medical events and temporal relations from Chinese electronic medical records,but they are far from deep enough,and mainly focus on the processing of English electronic medical records,while there are few studies on Chinese electronic medical records,which leads to the task of event and temporal relationship extraction of Chinese electronic medical records still needs further research.In this thesis,we study the extraction technology of medical events and temporal relations for Chinese electronic medical records.The specific research contents include:(1)This thesis constructs a corpus of temporal relationship of medical events based on Chinese electronic medical records to solve the problem of the lack of temporal relationship corpus of Chinese medical events.First of all,this thesis draws on the labeling norms and experience adopted by the i2b2-2012 temporal relationship corpus and THYMPE corpus,and designs the labeling norms of the Chinese medical event temporal relationship corpus according to the writing characteristics of Chinese electronic medical records.Secondly,according to the labeling specification,this thesis develops a labeling tool to realize the standardization of labeling and results.Finally,with the guidance of relevant personnel in the medical field and the mastery of the labeling norms by the labelers,the construction of a chronological relationship corpus of medical events for Chinese electronic medical records was completed.(2)This thesis proposes a Chinese medical event extraction method based on event frequency distribution ratio and document consistency to solve the problem of not considering the distribution characteristics of medical events and ignoring the document consistency distribution among medical events in each document.Firstly,domain adaptation on the Chinese pre-trained language model BERT is performed based on a large amount of Chinese electronic medical record texts.Secondly,on the basis of the basic features,according to the distribution information of medical events in electronic medical records,the event frequency distribution ratio is designed to select different event information as extended features.Finally,this thesis also uses document consistency information to enhance the extraction accuracy of medical events.Experiments were carried out on the Chinese CED corpus,and the F1 value of the method in this thesis reached 92.43%.At the same time,the results of the ablation experiment proved the effectiveness of the method proposed in this thesis.(3)This thesis proposes a medical event temporal relation extraction method,which integrates a bidirectional hierarchical labeling framework and global reasoning,to solve the problems of error propagation and local constraint inconsistency caused by unidirectional models.Firstly,based on the existing research experience,this thesis designs a two-way hierarchical labeling framework to make up for the problems in extracting “event pairs” by the one-way labeling framework,such as small quantity,poor quality and error propagation.Secondly,this thesis combines the dependence of event relations with the symmetry and transitivity of temporal relations,and designs global inference constraints to solve the problem of inconsistency in local constraints.Finally,the extraction of event timing relationship triples was completed on the self-constructed corpus,and the F1 value of 93% was achieved.The effectiveness of the bidirectional hierarchical labeling framework and the global reasoning constraints is verified by ablation experiments.