Research And Application Of Disease Prediction Based On Electronic Medical Records

Electronic medical records refer to the electronic notes of the patient's consultation information.They record all the information of the patient from the time of admission to discharge,and they can provide inquiry and decision support for the patient or doctor.The electronic medical records generally include image and text information,and the text information exists in the form of semi-structured text or free text.The medical data research based on the electronic medical records is of great significance.This thesis aims at the electronic medical records text,the main research contents are as follows:(1)De-identification of electronic medical records.The electronic medical records contain many private information.If the identified information is compromised,it will cause harm to the patient.Therefore,the private information must be identified and processed.First,the text is cleaned and normalized,and the basic structure and representation method are generated.A deep conditional random field model with boundary features is proposed and the optimal feature set is selected.The text is expressed as a form of word vectors and trained as the input of neural network.BRBiRNN and BR-BiLSTM-CRF named entity recognition models based on block representation are proposed.Experimental results show that the F value is higher than the traditional method.(2)Aiming at the hypertensive disorders in pregnancy,we propose a recurrent neural network prediction model based on a feature mixed method.Hypertensive disorders in pregnancy is a disease that only happens to pregnant women.The presence of the disease affects the health of pregnant women and fetuses.It is important to detect whether pregnant women are sick in order to ensure mother and child are healthy.In this thesis,after preprocessing the obstetrics and gynecology electronic medical records,a feature mixed method is proposed.After the word vectors and the part of speech vectors are stitched to obtain the mixed feature,the model can achieve better performance without any other professional medical experience.The word feature and mixed feature are used for comparison experiments.Experiments show that the use of mixed features is approximately 2% better than using a single word feature.(3)A TQ-LSTM prediction model of hypertensive disorders in pregnancy based on text quantification is proposed.Firstly,the text is expressed quantitatively by information extraction,physical parameters related to pregnant women's physical state are extracted as feature vectors,missing data are complemented.At the same time,compared with the general RNN model,the experimental results show that the TQLSTM model achieves prediction of hypertensive disorders in pregnancy and the accuracy is higher than the general RNN model.