Research On EHR Data Based Sepsis Prediction Model For ICU Patients

As a common disease in ICU,sepsis is a huge threat to patients and medical system because its pathogenesis and treatment methods have not been found.In order to predict sepsis in ICU and reduce the mortality of patients and the pressure of medical system,the research of sepsis prediction based on artificial intelligence has gradually attracted attention.Thesis focuses on the research of ICU sepsis prediction model based on EHR data,focusing on the extraction and transformation of electronic health record data,the research of prediction model,the improvement of model performance and the improvement of model generalization.The main research content is divided into three parts.The first part is the extraction and transformation of electronic health record data.In order to transform the original data to meet the requirements of machine learning for data,and to maintain the generalization of the transformation method,thesis proposes a general data transformation method for electronic health records.The data in the electronic health record system is divided into three categories and different methods of data processing,screening and transformation are adopted.The most useful features are retained while taking into account the generalization,which can be applied between different data record systems.The second part is the construction of prediction model based on graph neural network.Based on the characteristics of electronic health record data as time series data,thesis first constructs the most common neural network model to predict sepsis.In order to improve the accuracy of the model,thesis uses the graph convolution neural network for feature extraction,modifies the structure of the model,and uses the weighted sum of the loss at the graph level and the node level to calculate the loss.Compared with the existing neural network model,the AUROC of the proposed model is improved by 4–6%on the MIMIC dataset and about 3% on the Physionet Challenge dataset.The third part is the construction of the prediction model based on XGBoost.Due to the excellent performance of XGBoost,thesis proposes a prediction model for time series data based on XGBoost framework,which has achieved the best performance in accuracy and speed.The improvement of the model is mainly reflected in three parts.1)A new feature selection method is proposed to reduce the number of features and the impact on performance.2)A new feature construction method is proposed.This method constructs time series features for dynamic test features and statistical counting features for discrete features.Compared with no feature construction or other feature construction methods,the feature construction method in thesis is greatly improved.3)A new evaluation function is proposed,and the first and second steps are given,which improves the accuracy of the model to a certain extent.Combined with these improvements,the AUROC of the model is improved by about 5% on the MIMIC dataset compared with the existing model,and by about 3% on the Physionet Challenge dataset.Finally,the sepsis data set construction method and model construction method proposed in thesis do not involve the processing of specific features.They can be uniformly processed according to the classification of data,and can be migrated in different electronic health records,with strong generalization.