Toward Chinese Electronic Medical Records: Research And Implementation Of Quantitative Information Extraction Method

With the development of medical digitalization,the data volume of EMR in medical system is increasing in a geometric level.As the quantifiable data of patients' condition,medical quantitative information is widely distributed in electronic medical records in an unstructured way.Due to the complexity of medical information representation and the absence of research on medical quantitative information extraction,how to extract quantitative information from unstructured electronic medical records efficiently and accurately has become an urgent problem.Therefore,this research focuses on the extraction method of CQI in Chinese EMR,and the main achievements of this paper are as follows:1)In this paper,1359 cases of Chinese electronic medical records in burn department of a third-class hospital were systematically and manually annotated.We proposed a rule-based method based on self-building dictionary and a pattern-learning method based on semantic role for quantitative information extraction.In addition,we built and implemented a series of typical sequential labeling models for Chinese medical quantitative information extraction.Moreover,in order to provide a comparative analysis and research for the extraction of medical quantitative information,we carried out a comprehensive horizontal comparison to evaluate the extraction methods with the same labeled medical data set.2)This paper proposed a BiLSTM-CRF optimization model incorporated with external features for medical quantitative information recognition.By using the artificially integrated burn medical dictionary,we obtained the dictionary features in the longest matching way.By treated the numerical information as the center target,we achieved the location features of each characters.For exploring the influence of different feature fusion methods,we integrated those two kinds of external features into the representation of the model,including input layer splicing method and hidden layer splicing method.The experimental results under the same test set showed that the model incorporated with external features is greatly improved compared to the baseline model,and the extended model that uses input layer splicing method performed the best among all models.It is proved that the external features proposed in this paper can effectively improve the ability of deep learning model for capturing the characteristics of medical quantitative information on a limited data set.3)For further entity-quantity association,this paper proposed an automatic association method based on Feature Engineering.By extracting the quantitative information structural features of sentence level,four kinds of structural features are formed.Besides,utilizing the non-parameter outlier detection method to train the standard data set,we obtained the statistical-based numerical attribute features.Finally,the feature set was screened by correlation testing and feature ablation.Using the random forest as the classifier,we realized the automatic association between entities and quantities.The evaluation result of all tested mainstream classifiers is higher than 87%,and the random forest classifier obtains the highest correct rate(95.50%),proving the efficiency of entity-quantity association method proposed in this paper,which solved the problem of labor cost and low generalization of rule-based method.Finally,the associated medical quantity information was standardized by semantic representation and the structured medical quantity information was formed.