Research On Dynamic Prediction Of Blood Glucose And Abnormal Early Warning Driven By Knowledge And Data

Diabetes Mellitus is one of the fastest gro wing chronic diseases in the world in the 21st century,which does great harm to human health,society and economic development.Blood glucose regulation is an extremely complicated process because it is affected by many factors.In order to better control blood glucose in the normal range,so as to more effectively reduce the fatality rate of diabetes,improve the quality of life in patients with diabetes,how to accurately estimate the future blood glucose and to detect possible abnormalities in advance has become an important subject for many medical workers and researchers.First,under the clinical CGMS data,the states are divided based on the fact that there are multiple states of the patient,and the key decision variables are extracted in combination with the clinician’s experience-knowledge.The decision tree model is used to mine effective classification rules for state from historical blood glucose data.The clinical experience-knowledge is re-expressed to build a rule base.The experiment judges the states after half an hour and comparatively analyzes the dual-decision attributes and the four-decision attributes.The results show that the extract classification rules are effective,which lays the foundation for the establishing a multistates blood glucose prediction model.Secondly,combined with knowledge of various physiological states classification rules,a multi-state blood glucose prediction model based on Gaussian process regression is proposed.the appropriate kernel function is selected according to the analysis of the characteristics of CGMS data in each state.The Gaussian process regression method is applied to model for each state separately,and finally an overall model is formed.The experiment uniformly predicts the blood glucose in the next 30 minutes and comparatively analyzes the situation without distinguishing various states.The results verify that the multi-state blood glucose prediction model has higher accuracy and comprehensiveness and also provides a guarantee for the early warning.Finally,a multi-level clinical early warning mechanism for blood glucose abnormalities was designed.According to the different states of patients with different sensitivities to internal hormones and external stimulus,variable threshold strategies under different states were proposed.And according to the hazard levels of abnormal blood glucose in clinical manifestations,a hierarchical alarm strategy under different hazard levels was proposed.The experimental results show that the variable threshold strategy based on different states can slightly increase the false positives and greatly reduce the false negatives,reducing the risk of dangerous patients.The hierarchical alarm strategy based on different hazard levels makes the blood glucose abnormal warning system more patient-friendly in clinical application.