A Machine Learning Prediction Model For Electronic Health Records Based On Prior Medical Knowledge

With the help of artificial intelligence,Internet of things and other technologies,intelligent medical care aims to realize personalized and mobile medical care.In recent years,predicting the risk of potential diseases through electronic health records has attracted extensive attention in the field of artificial intelligence.With the development of deep learning algorithm,compared with the traditional machine learning model,the method based on deep learning can achieve better results in risk prediction tasks.However,existing work takes little account of prior medical knowledge,such as the relationship between disease and associated risk factors.Patient Electronic Health Records(EHR)data includes a sequence of visits over time,in which each visit contains multiple medical codes,including diagnostic,medication and procedure codes.Therefore,how to deal with continuous and high-dimensional EHR data and integrate the existing medical rules into the existing risk prediction model to improve the prediction accuracy has become two major problems faced by intelligent healthcare.In order to cope with the above challenges,this thesis proposed following solutions:First,for high-dimensional EHR data,this thesis adopts bidirectional cyclic neural network to remember all information of past and future visits,and introduces three attention mechanisms to measure the relationship between different visits for prediction.Experimental results show that the proposed method can effectively process the high-dimensional data with low-dimensional.Second,in order to improve the accuracy of risk prediction,this thesis proposes a new and general framework called risk prediction task PRIME,which can successfully apply the discrete PRIor MEdical knowledge to the prediction model using posterior regularization technology.Thirdly,the neural network was used to automatically set the boundary value of each priori medical knowledge to simulate the required distributionof the patient's target disease.The proposed RIME framework of risk prediction based on machine learning can automatically learn the importance of different prior knowledge through log-linear model.Experimental results show that compared with traditional machine learning and neural network methods,the proposed model can improve the accuracy of risk prediction in three real medical data sets.