Risk Prediction With Electronic Health Record:A Deep Learning Approach

Data driven healthcare aims to use large-scale data,through data analysis,machine learning and other methods to achieve disease risk prediction.In order to provide the best and most personalized care,electronic health record containing a lot of valuable information will be used in the data-driven medical revolution.However,a single admission in EHR contains multiple diagnoses of different size which makes the admission information difficult to express,and each admission has irregular time.In addition,the disease has long-term dependence.These are the challenges to realize disease risk prediction.With the development of knowledge and technology in the field of deep learning,it is of practical significance to study the method of disease risk prediction based on electronic health record using deep learning technology to provide reference for doctors' diagnosis,so as to improve the timeliness and accuracy of clinical diagnosis and reduce medical costs.However,no much research works can be found.In order to solve the above challenges,the major contributions of this thesis can be described as follows:1.The information of EHR is difficult to be expressed because that the hospital information recorded in EHR is very sparse.Based on the word embedding method in natural language processing(NLP),a specific word vector is developed through the coding representation of diagnosis,and then a hospital admission is represented by another more detailed word vector.In this way,the hospital admission with different size is embedded.A continuous vector space will be used as the input of the prediction model.2.In order to solve the temporality problem of EHR's information,a Long Short-Term Memory model is proposed.For irregular timing,the proposed model has a different LSTM forgetting gate by modifying the forgetting mechanism.The experimental results show that the proposed model is effective compared with Markov and cyclic neural networks.3.Chronic disease is a long-term disease.However,LSTM is not explicitly designed to predict a long-term disease.Based on the LSTM model proposed earlier,a hybrid depth neural network(LSTM + CNN)model is proposed to integrate diagnosis information through the introduction of convolution neural network.The validity of the model is shown by comparing it with Naive Bayes,Random forest and other deep learning methods.4.The design and implementation of a disease risk prediction software based on deep learning methods.The software includes electronic patient record information management,prediction model and visualization.The software can make disease risk prediction more accurate and timely,and provide helpful information for doctors' diagnosis.