Research On Online Medical Information Extraction Based On Deep Learning

"Internet + medical treatment" is changing the life style of people with the development of internet technology and the convert of public health self-management consciousness.Nowadays,there appears a lot of online medical communities,medical health information websites and health care APP for all kinds of user at home and abroad.Those kind of online medical health platforms mainly provide the service like health knowledge,disease information,drug information,medical health news,disease inquiry and so on.Patients,medical personnel,medical researchers and other users can descript,share,and consult about disease information about drugs,treatment process,treatment,new medical knowledge on online medical platforms.The well-known domestic online medical platforms are HaoDF,ChunYuYiSheng,.CMPDB,etc.There are a large number of active users on these platforms.Online medical platforms contain huge amounts of medical and health related data,the data contains abundant medical value.But such online medical text most is unstructured text.Medical health entity extraction,entity relationship extraction and medical entity attributes extraction is usually the first step to have a further mining and utilization of these huge amount of unstructured online medical text.Nowadays,information extraction mainly focus on social media text,news text and other daily field,and mainly studies on person named entity recognition,place named entity recognition,named entity recognition and so on.In the field of medical health,information ext:raction mainly focus on professional text such as electronic medical records,discharge summary,clinical text.There are few study on extraction of online medical information.The main used method in traditional information extraction is the machine learning methods such as Hidden Markov Model,Support Vector Machine,Conditional Random Field.The traditional information extraction methods are heavily dependent on manual extraction features.Manual extraction features take not only a lot of time and cost,but also the extracted features are also limited.Deep learning methods can effectively solve this problem.It can complete the automatic extraction and presentation of features through the deep neural network.And it has been proved that the information extraction based on the deep learning is superior to the traditional information extraction method when dealing with massive amounts of data.This paper analyzes the sub-language characters of online medical text.Based on these sub-language characters to build mixed deep learning models to extract the medical entity,medical entity relationship,medical entity attributes.This paper proves the validity and usefulness of the mixed models by comparative experiments and examples.The main research content mainly includes the following five aspects.(1)Revealing the sub-language characteristics of online medical text,and on this basis,the paper constructs the online medical information extraction framework based on depth learning.The statistical analysis method was used to analyze the sub-language characteristics of online medical text and clinical text.This paper studies the sub-language features of online medical text from the perspective of the content contained in the text,the frequency of use of the word class,and the main semantic category of the text.Based on the sub-language characteristics of online medical text,this paper analyzes the existing deep learning models and chooses the applicable models to construct the online medical information extraction framework.(2)Constructing a hybrid depth learning model DNN-LSTM to extract medical entities.The medical entity type and the medical entity extraction goal in the online medical information extraction task are defined based on medical entity identification task of the electronic medical record in i2b2 2010.According to the problem description of medical entity extraction task,a hybrid depth learning model CNN-BLSTM framework is constructed.Based on the online medical text data processing process,the medical entity extraction process based on CNN-BLSTM was discussed from five stages:data preprocessing,feature selection of medical entity extraction,feature embedding of medical entity extraction,BLSTM layer and the output of tag sequence.Three groups of experimental results show that the hybrid depth learning model CNN-BLSTM is better than CNN model and BLSTM model on medical entity extraction in online medical text.(3)Constructing a hybrid depth learning model LSTM-CNN to extract medical entity relationship.The medical entity relationship type and the medical entity relationship extraction goal in the online medical information extraction task are defined based on medical entity relationship identification task of the electronic medical record in i2b22010.According to the problem description of medical entity relationship extraction task,a hybrid depth learning model BLSTM-CNN framework is constructed.In the BLSTM-CNN model framework,firstly,the output characteristics of each word of the sentence are integrated through the BLSTM layer,and then complete the semantic learning of the whole sentence.Secondly,according to the position of the two medical entities in the sentence,the sentence feature is divided into three parts.The CNN model is used to convolution and pool the three parts.The feature vector of the sentence is extracted by the CNN.Finally,the sentence feature vector is sent into the softmax classifier for medical entity classification.The experimental results show that the hybrid deep learning model BLSTM-CNN constructed in this paper is better than BLSTM model and CNN model on medical entity relationship extraction in online medical text.(4)This paper studies the application of two kinds of mixed depth learning models in medical entity attribute extraction task.Medical entity attribute extraction can be seen as a sequence label problem,can also be seen as a classification problem.There are some differences between the features selection of medical entity attribute extraction and the features selection of medical entity extraction and medical entity relationship extraction.The two kinds of mixed depth learning models were used to extract the medical entity attributes respectively.The futures are reanalyzed and selected in the extraction of medical entity attributes.The experimental results show that the mixed depth learning model CNN-BLSTM has a better result on the extraction of medical entity attributes in online medical text.(5)To explore the application of online medical information extraction in disease association detection.In this paper,the possible application fields of online medical information extraction results are summarized,and disease association detection was selected for specific application demonstration.Depending on the type of PIP relationship between the disease entities,and the temporal attributes of the disease and patients,to identify co-occurrence and causal relationship between diseases.Finally,using the medical health guidelines and information to verify the relationship between diseases.