The Research Of Clinical Medical Prediction Model Based On Deep Neural Network

With the development and application of medical information technology,various health-related physiological indicators of patients are recorded in electronic medical records.By analyzing and mining electronic health records,the combination of different detection indicators in electronic health records can be extracted.Auxiliary analysis of health status to provide further medical predictions.The use of deep neural networks can more accurately predict the patient’s current physical condition and provide personalized diagnostic prediction services.Therefore,it is important to study clinical medical prediction models based on deep neural networks.This thesis focuses on the method of clinical medical prediction model based on deep neural network.The main work is as follows:(1)Analyzing the ICU clinical medical database and create four predictive task data sets.In order to meet the requirements of rich data and wide applicability,the open MIMIC-Ⅲ database is selected as the data source,and four current data sets that are meaningful and urgently needed to address clinical predictive tasks(decompensation prediction,in-hospital-mortality prediction,length of say,phenotyping)are produced.(2)Applying deep neural network to clinical medical prediction,three clinical medical prediction models based on deep neural network are proposed and compared experimental analysis.Based on the clinical medical prediction model of PNN,considering the combination of all physiological detection indicators of patients,it can avoid ignoring some combinations of indicators that are important for determining the physical condition of patients;based on the clinical and medical prediction model of Wide and Deep,the lower order of patient characteristics is considered.Cross-question can remember physiological indicators that are important for determining the current physical condition of patients;DeepFM-based clinical medical prediction model overcomes the sparseness of electronic health record data,can automatically extract physiological detection index combination features,and efficiently construct cross feature.Compared with the benchmark evaluation model experiment results,Wide and Deep and DeepFM improved the effect of three prediction tasks(decompensation prediction,in-hospital-mortality prediction,phenotyping).(3)Improve the Wide and Deep model and propose a clinical medical prediction model based on Wide and Din.The model uses a local activation mechanism to achieve better prediction results through the joint training of Wide and Din.Compared with the experimental results of the benchmark evaluation model,the model significantly improves the prediction effect of three tasks(decompensation prediction,in-hospital-mortality prediction,phenotyping).