Deep Neural Network Prediction Model Based On Electronic Health Records

Electronic Health Records are valuable research resources because of the abundant information they contain.The mining of EHRs can improve the efficiency of medical diagnosis and improve the timeliness,accuracy of clinical diagnosis and reduce the cost of medical activities.The EHRs processing is one of the basic of Internet medical.In this thesis,we use deep learning method,build deep neural networks predict model.With some traditional machine learning methods,this thesis introduces the predict model of medical records de-identification,fetal weight estimation and disease predict classification.The main contributions of this thesis are as following:To tackle the problem that EHRs cannot be easily accessed by researchers or organizations for a large amount of protected health information(PHI)exists in EHRs,we proposed the Text Skeleton based recurrent neural network de-identification framework.Text skeleton is the general structure of a medical record,which can help neural networks to learn better.We evaluated our method on three datasets involving two English datasets from i2b2 de-identification challenge and a Chinese dataset we created.Empirical results show that the text skeleton based method we proposed can help the network to recognize protected health information.The comparison between our method with state-of-the-art frameworks indicates that our method achieves high performance on the problem of medical record de-identification.Specifically,the performance on two different i2b2 datasets as well as the Chinese dataset demonstrated an F-score of about 0.98 consistently.For the challenge of fetal weight prediction,we proposed a deep neural network structure for building fetal weight prediction model.Traditional fetal weight prediction models are based on medical knowledge and relay on feature selection,which is leading to the hard generalize of the model building process.What is more,we introduce the process which extracting parameters from electronic health records and the filling strategies for missing value.Empirical results show the deep neural network based prediction model outperforms traditional methods with reduce the error by 8.9 percent.We also introduce a w-KNN method for disease prediction model in the thesis.The model first transit the medical records to structured data,and map the short text to symptom phrase.An important advantage of w-KNN model is it can find out the similar samples from the knowledge base.The similar sample can help doctor make decision,and give patient better understand of their situation.Empirical results show that the w-KNN model can predict disease accurately.