Research On Relation Extraction In Electronic Medical Records

As the EMR contains so much knowledge in the medical fields and has a hugequantity of corpus, using the informatics resources to help diagnose and build uppersonal health model can be a very important part in the development of MedicalInformatics. The discharges and progresses of EMR are unstructured text whichrecords the medical progresses, and they contain many professional words. So themethods of information extraction can be the first step in the work of knowledgemining in EMRs. And the concept extraction and relation extraction can be thekernel part. Getting labeled data can be very difficult because of the need ofprofessional knowledge in the EMRs. The conference of i2b2improves the missionof sharing corpus and the researches on the EMRs.This paper focuses on the mission of relation extraction in the EMR, the work ofresearch includes four parts:(1) Introduce the mission of relation extraction and its evaluation, and the corpusdata used in the paper.(2) Give a brief introduction to the basic method of relation extraction and selectbasic features to build a classifier by CRF to convert the mission into aclassification problem as a baseline experiment. After reviewing the badcases in the classification, this paper made a research on feature selection onthe data of EMRs.(3) Introduce the feature improving method by deep learning. Find a new way ofrepresenting the data to get more discriminative features. This experimentuses the deep sparse auto encoder to make layer-wise feature abstracting withan appropriate parameter. The result showed that there was an improvementin the recall of the bad cases and help increase the F1value by1.5%to get86.1%. The result showed that representing the data by deep learning canfind discriminative features in the EMRs.(4) Another way of improving the result was built on the relations betweenwords, this paper first merged the words which have similar meaning bystatistics and dictionaries, and then found a way of statistics to extractdiscriminative features. This method improved much in time consumecompared to the way of deep learning. This method got improvement in theresult after selecting appropriate parameters in the experiment. Finally themethod help increase the F1value by2.3%to get86.9%. The result provedthat statistics method after merging similar words can be an effective way of feature extraction in the mission of relation extraction in EMRs.In a conclusion, this article focuses on the mission of relation extraction inEMRs, and the methods of finding method to get more discriminative features b ydeep learning and statistics show better results than basic method. So the methods inthe paper can be effective methods in the mission of relation extraction in EMRs.