Data Mining And Application Of Intensive Care Database

The data used in this study comes from the intensive care medical information database(Multiparameter Intelligent Monitoring in Intensive Care I,MIMIC database),which consists of five data sources or modules: hospital database,hospital supplier orders,online resources,Carevue database(CV)and Metavision database(MV)Integration,which leads to storage differences in some information,and brings great difficulties to data cleaning,data sorting,data unification and analysis.Based on the MIMIC database,we perform fuzzy matching based on keywords such as the name and unit of the indicator,and develop context-specific to define the criteria and rules for indicator extraction.The distributions of the obtained indicators are analyzed,and the data is subject to preprocessing.After data preprocessing and screening,the relevant clinical information for patients that meet the diagnostic criteria are obtained.These indicators are subject to further deduplication of highly correlated parameters,followed by normalization or standardization.Lasso is used to find independent risk factors that affect patient mortality.We use grid search to combine different parameters of the algorithm and establish multiple decision tree models and random forest models.Finally,the Kappa score of each model is calculated,and the corresponding receiver operating characteristic curve(ROC curve)is plotted to evaluate and select the optimal model.The model for research results and applications includes the following two aspects:(1)We evaluate the relationship between the sequential organ failure score(SOFA score)and mortality of all diseases entering the intensive care unit in the database;(2)We use a comprehensive clinical diagnostic criteria to extract information from the database for patients in hyperglycemia crisis admitted to the hospital for diabetes for the first time,and used acute complications of diabetes as a case study to evaluate our MIMIC database mining method.Finally,we have established a data mining workflow for analyzing and predicting the mortality risk from the MIMIC database.The developed interpreter can fetch and integrate corresponding data from different data sources,providing a good foundation for further modeling.The data mining process and interpreter used can be applied to the study of patients with acute complications of diabetes.Taking the acute complications of diabetes as a case study,we demonstrate that our model may be applied to predict the mortality risk from related diseases,and the results may be useful to provide guidiance for clinical diagnosis and treatment.