Factors Contributing To The Error In Glucose Prediction And Evaluation Of Predictive Features Based On The GM(1,1) Model

BackgroundDiabetes mellitus is one of the global problems that jeopardize human health,and its early diagnosis and rational treatment have a profound impact on improving patient outcomes and reducing the burden on families and society.Blood glucose monitoring plays a crucial role in the diagnosis and treatment of diabetes.With the maturation of blood glucose monitoring technology,using machine learning algorithms to predict blood glucose becomes one of the research hotspots.Machine learning algorithms include linear algorithms,nonlinear algorithms,etc,among which the GM(1,1)model belongs to one of the nonlinear algorithms and has the advantages of few required samples,simple operation,and good accuracy.Therefore,this study evaluates the influencing factors of the glycemic prediction error and predictive features of the GM(1,1)model to guide clinical decision-making.Objective1.To explore the main fluctuating factors affecting the prediction error when predicting blood glucose by the GM(1,1)model.2.To compare the predictive effect of the GM(1,1)model for different ranges of blood glucose and in turn to evaluate the predictive characteristics of the model.MethodsIn this study,blood glucose information was retrospective collected from 142T2 DM patients who were hospitalized in the Endocrinology Department of Henan Provincial People’s hospital.The GM(1,1)model was applied to predict blood glucose after 5 min,10 min,15 min,20 min,25 min and 30 min,and the error between measured and predicted values was analyzed to evaluate the predictive performance of the model.To search for major fluctuating factors affecting the prediction error,Pearson’s correlation analysis and the multiple linear regression analysis equations were performed for prediction error(MAE)and the following indexes: standard deviation of blood glucose(SDBG),mean blood glucose(MBG),large amplitude of glucose excursions(LAGE),mean amplitude of blood glucose excursions(MAGE),number of effective blood glucose excursions(NGE),standard deviation of amplitude of glucose excursions(SAGE),coefficient of variation of AGE(AGE-CV),total effective glucose excursions(TEGE),total blood glucose excursions(TGE).To evaluate the predictive characteristics of the GM(1,1)model,the patients were divided into two groups,50 in the observation group and 92 in the experimental group.The observation group data were used to derive the criteria for dividing the different blood glucose groups,and the experimental group data were used to evaluate the predictive ability of the GM(1,1)model in different blood glucose ranges.Results1.Pearson’s correlation analysis indicated TEGE,TGE(P < 0.001)were highly correlated with prediction error,SDBG,LAGE and MAGE(P < 0.001)were strongly correlated,SAGE(P < 0.001)was moderately correlated,MBG(P < 0.001)was weakly correlated,and AGE-CV,NGE(P > 0.05)were not correlated.After building multiple linear regression equations,TGE,LAGE,MAGE and NGE were found to be independently associated with prediction error,among which the coefficient of MAGE was the largest.2.When predicted 5 min in advance,the GM(1,1)model ’s prediction efficiency to blood glucose at levels 3.4 ～ 13.3mmol/L,more than 13.3mmol/L,less than3.4mmol/L gradually decreased;When predicted 15 min in advance,the model’s prediction efficiency sequentially decreased for blood glucose of 3.4 ～ 14.4mmol/L,more than 14.4mmol/l,less than 3.4mmol/L;When predicted 30 min in advance,the model’s prediction efficiency at levels 3.2 ～ 13.8mmol/L,more than 13.8mmol/L,less than 3.2mmol/L gradually decreased.Conclusions1.MAGE,LAGE,TGE,and NGE are the main fluctuating indexes that affect the accuracy of glycemic prediction of the GM(1,1)model,among which MAGE has the highest influence.2.When it was predicted 5min,15 min,30min in advance,the GM(1,1)model had the best prediction effect on the blood glucose data which was at 3.4 ～ 13.3mmol/L.