Medical Named Entity And Assertion Recognition In Chinese EMRs With Deep Learning

In recent years,with people’s attention to health,intelligent medicine has become a research hotspot,among which the information extraction of electronic medical records is the basic work.Electronic medical record refers to the text information generated by medical information system in the process of medical activities,and it is a kind of digital information of clinical records.Through the analysis of electronic medical records,a large number of medical information closely related to patients can be excavated,which can play a great role in clinical decision-making.Natural language processing technology can help us to extract text information from electronic medical records,such as named entity recognition,medicalassertionclassification and so on.Entity recognition aims at identifying the boundary and category of different entities in medical records,and medical assertion classification aims at identifying the modification relationship between specific entities and patients.However,compared with other texts,there are many problems in Chinese electronic medical records,such as irregular writing,more professional terms,frequent abbreviations for special characters,incomplete sentence structure and so on.At the same time,due to the privacy of patients,the available electronic medical record data is limited.All these characteristics make it more difficult to format electronic medical records.Therefore,in order to extract the electronic medical record information better,this paper studies the entity and modification recognition of Chinese electronic medical record and its joint extraction technology on the independently labeled data set by means of in-depth learning method.(1)Entity recognition based on knowledge attention mechanism enhancement.Although previous studies have achieved good performance,they neglect the external medical knowledge which can provide abundant entity definition and entity boundary information.Therefore,we propose a CNN-BLSTM-CRF model based on medical knowledge attention mechanism enhancement.By using the attention mechanism,the definition of medical entity and the boundary information of entity in medical dictionary are coded to enhance the performance of the neural network model.Before encoding text information with BLSTM,the character-level representation of text is extracted in advance by CNN as supplementary information of text information,which effectively solves the problems of irregular writing and frequent special words in the electronic medical records mentioned above.(2)Medical assertion classificationbaseds on CNN-GRU neural network.GRU network is used to encode the text information of electronic medical records,and Softmax is used to decode it.Among them,GRU network is a variant of recurrent neural network(RNN).Compared with GRU,it solves the problem of long-distance dependence better and calculates in a simpler way.At the same time,we still use CNN network to pre-extract character level representation to solve the problems of irregular writing and frequent special characters.(3)Joint extraction of entities and assertion.In joint extraction,we propose a joint extraction method based on multi-label scheme,and use BLSTM-LSTM model to identify.Compared with the traditional serial extraction method,we transform two tasks into one task through tagging scheme.Instead of extracting entities first and then extracting entity modification categories,we use an end-to-end model to extract entities step by step through multiple tags.The experimental results show that our method is superior to previous methods on the same data set.