Blood Glucose Prediction And Hypoglycemia Warning Evaluation Based On LSTM-GRU Model

Research backgroud:The latest report released by the International Diabetes Federation(IDF)in 2019 shows that the incidence of diabetes mellitus is increasing year by year,From 285 million in 2009 to 537 million in 2021,it is expected to increase to 578 million in 2030 and 693 million in 2045,becoming the third chronic non-communicable disease that seriously affects human health after cardiovascular and cerebrovascular and tumor diseases.A series of complications of diabetes mellitus not only reduce the quality of life of patients,but also bring greater economic pressure to patients and their families.Therefore,diabetic mellitus patients need self-monitoring of blood glucose(SMBG)to control blood glucose levels within the normal range,which is essential to reduce diabetes-related complications.At present,blood glucose levels monitoring technology is becoming more and more mature,and the use of machine learning algorithms to predict blood glucose levels has become one of the research hotspots.Machine learning algorithms mainly include linear algorithms,nonlinear algorithms,etc.Among them,LSTM and GRU are nonlinear algorithms.LSTM has high prediction accuracy,but its complex internal structure reduces the training speed of the model,while GRU has the characteristics of short training time.Therefore,some scholars try to combine LSTM and GRU for sentiment analysis,traffic flow prediction and other fields,which improves the model’s accuracy and relatively reduces the model’s running time.Due to the working principles of the LSTM and GRU are similar and these two models are deficient in each other,this paper attempts to combine the two models to establish the LSTM-GRU model for blood glucose levels prediction,and compare its prediction performance with the LSTM,GRU and GM(1,1)models at different prediction horizons.On this basis,the LSTM-GRU model is used for hypoglycemia warning,and the main factors affecting hypoglycemia warning performance of the model are explored,which has specific clinical value in reducing the occurrence of hypoglycemia events.Objective:1.Compare the blood glucose levels prediction performance of the LSTM,GRU,GM(1,1),and LSTM-GRU models at different prediction times.2.To explore the hypoglycemia warning performance of the LSTM-GRU model and its main influencing factors.Methods:1.A retrospective analysis of the blood glucose levels information of 100 DM patients who received subcutaneous insulin pump therapy in the Endocrinology Department of Henan Provincial People’s Hospital from March 2017 to December 2017.Due to patients and doctors need at least 15 min to adjust the treatment plan,and the blood glucose levels prediction error changes every 5 min is small.Therefore,this paper takes 15 min,30min,45 min,and 60 min as the time cut-off points,the LSTM,GRU,GM(1,1),and LSTM-GRU models were used to predict blood glucose levels,and the predicted blood glucose levels of each model at different times were obtained respectively.Calculate the prediction error between each predicted blood glucose level and the continuous glucose monitoring system(CGMS)measured blood glucose level.Correlation analysis and Clark error grid analysis were carried out on the measured blood glucose level of CGMS and the predicted blood glucose level of the model,and the training time of the four models was compared to evaluate the prediction performance of the model.Since RMSE can avoid dimensional problems,it is more sensitive to outliers.Therefore,this study uses Root Mean Square Error(RMSE)as the evaluation index,and repeated measures ANOVA was used to compare whether the prediction errors of the four models under different times were statistically different.If there is a statistical difference,compare the error sizes of the four models under the time of 15 min,30min,45 min,and 60 min,and combine the statistical analysis results and the error size to get the best blood glucose level prediction model among the four models.If there is no statistical difference,it means that the prediction errors of the four models at different times are not different.2.Taking 3.9mmol/l as the hypoglycemia threshold,according to the blood glucose level predicted in the previous part of the LSTM-GRU model,a hypoglycemia warning is immediately issued for the blood glucose level predicted by the model below 3.9mmol/l,and different hypoglycemia intervals are set for partitioning warning.According to the relationship between whether the blood glucose level predicted by the model is hypoglycemia and whether the measured level of CGMS is hypoglycemia,the hypoglycemia warning sensitivity,false-positive rate,false-negative rate,specificity,and accuracy of the model are calculated.Sensitivity is one of the main indicators to evaluate the model’s correct early warning of hypoglycemia events.Repeated measures ANOVA was used to compare the sensitivity of 15 min,30min,45 min,and 60 min.The main factors affecting the hypoglycemia warning performance of the LSTM-GRU model were explored using the ROC curve.Results:1.The predicted levels of the four models are significantly correlated with the measured levels(R>0.5,all P<0.001).The correlation coefficient(R=0.995)of the LSTM-GRU model does not decrease over forecast time,and is higher than the GM(1,1)model at the same time,while the R-values of LSTM and GRU gradually decreased after 30 min.Under the same prediction horizons,the proportion of the LSTM-GRU model in the zones A+B is higher than that of the LSTM,GRU and GM(1,1)models.GRU and the LSTM-GRU models have the shortest training time,followed by the LSTM model,and the GM(1,1)model has the longest training time.With the horizontal extension of the prediction time,the RMSE of the LSTM,GRU and GM(1,1)models gradually increased,and the RMSE of the LSTM-GRU model keep steady within 30-60 min.At different prediction times,the RMSE of the LSTM-GRU model,GM(1,1)model were statistically different from the RMSE of the LSTM model,GRU models(all P<0.001).the RMSE of the LSTM-GRU model and GM(1,1)model was statistical differences at different prediction times(all P<0.001).the RMSE of LSTM model and GRU model were not statistically different at different prediction times(all P>0.05).Under the same prediction time horizon,the RMSE of the LSTM,GRU and GM(1,1)models are higher than that of the LSTM-GRU model.Under the same prediction error,the prediction horizon of the LSTM-GRU model is longer than that of the LSTM,GRU and GM(1)Model.In summary,the LSTM-GRU model has the best predictive performance among the four models.2.There is no difference in the early warning sensitivity of the LSTM-GRU model at 30 min,45min,and 60 min.The hypoglycemia warning sensitivity and false-negative rate of the model stabilized after 30 min.With the extension of the prediction time,the false-positive rate,specificity,and accuracy of the model remained steady,showing good hypoglycemia warning performance,Therefore,60 min is selected as the time node for hypoglycemia warning.As the threshold increases,the sensitivity of the LSTM-GRU model increases and the specificity decreases,the accuracy of the hypoglycemia warning decreases as the forecast time increases.Conciusions:1.The prediction performance of the LSTM-GRU model at different times is better than that of the LSTM,GRU,and GM(1,1)models,and the prediction performance tends to be stable within 30-60 min.2.The LSTM-GRU model maintains a good hypoglycemia warning performance within 60 min,and the prediction time horizon and threshold are the main factors affecting hypoglycemia warning.