Research On Data-Driven Online Blood Glucose Prediction Method In Type 1 Diabetes

Type 1 diabetes,also known as insulin-dependent diabetes,relies more heavily on external insulin to control blood sugar levels than other types of diabetes.The online blood glucose prediction model is the basis of closed loop control of artificial pancreas.The model can be used to adjust insulin dose,close insulin pump management,set up high/low blood glucose warning and alarm,etc.Accurate blood glucose prediction model is the basis of reliable artificial pancreas system.Therefore,the high precision online blood glucose prediction method has become one of the key clinical problems in the field of blood glucose research.In this dissertation,the online blood glucose prediction method for type I diabetes was studied.The main research results are as follows:(1)The Variational Mode Decomposition and Gaussian Process Regression(VMD-GPR)method for online blood glucose prediction based on sliding time window was proposed to realize multi-step online blood glucose prediction.The edge effect of VMD is solved by using the method of data prediction extension.Then,the original blood glucose data is decomposed by VMD method to obtain multiple sub-modes.Compared with the original blood glucose data,the sub-modes are more stationary.According to the data characteristics of the decomposed sub-modes,the GPR method is used to forecast the.decomposed sub-modes,and the multi-step prediction is realized by the reconstruction method.Finally,the method of sliding time window is used to realize the online update of blood glucose data,so as to eliminate the data with weak correlation with the current moment and reduce data redundancy.The simulation results show that the prediction accuracy of VMD-GPR method is better than that of GPR method.(2)In the VMD-GPR model,how to get the best decomposition layer K and penalty factor a are very important to the model.The Gray Wolf Optimization(GWO)is used to optimize the VMD parameters,and the minimum local envelope entropy is used as the fitness function to construct the model based on GWO-VMD-GPR.The simulation results show that the GWOVMD-GPR model has better prediction effect than the VMD-GPR model.(3)To construct the online blood glucose prediction method based on blood glucose submodal division.The Sample Entropy method is used to divide the sub-modes into different complexity sub-modes,namely,the blood glucose fluctuation trend mode group and the detail mode group.For the trend mode group of blood glucose fluctuation,it has the characteristics of small complexity and strong regularity.Therefore,the GPR method is used to calculate the posterior probability after learning the joint probability distribution and then obtain the forecast value.The KELM method is used to directly learn the decision function for the detail modes with large complexity and weak regularity.Therefore,this dissertation integrates the GPR method and KELM method,and makes up for the shortcomings brought by using a single method to build the prediction model by combining the advantages of the two methods.Finally,the simulation experiment proves that the online blood glucose prediction model based on submodes partition has more advantages.