Study On Methods Of Blood Glucose Trend Prediction Based On Time Series Data

As one of the most common chronic diseases today,Diabetes can cause multi-organ complications in the late stages and can be life-threatening in severe cases.Currently,there is no clinical cures for Diabetes,and predicting blood glucose trends can help people who need to control their blood glucose levels to make timely responses to their current blood glucose state.Considering that the change of blood glucose is closely related to the characteristics of previous blood glucose level and dietary sugar intake,such characteristics data can help to provide more accurate and personalized prediction results of blood glucose trend for people who need to control their blood glucose.In this paper,three blood glucose trend prediction models were proposed from such characteristic data.The main work of this paper is as follows.(1)In this study,dynamic glycemic,heart rate,and diet-related characteristic data(including: Glycemic Index(GI),Glycemic Load(GL),and Carbohydrate(C)content)were collected from humans.In order to facilitate the construction of the model,the characteristic data were appropriately cleaned and converted into time series according to the data characteristics.(2)A method based on wavelet decomposition and Gaussian surface fitting was proposed for predicting blood glucose trends based on the GL of dietary intake and before meal and after meal glucose change data monitored by Continuous Glucose Monitoring(CGM).The wavelet decomposition method is used to decompose and reconstruct the original CGM data and smooth the curve,and a Gaussian surface was fitted to the blood glucose fluctuations under different GL data by intercepting the 1.5 h CGM data before and after the meal.This method can derive the blood glucose fluctuation before and after meal only using the dietary intake GL,and it showed better prediction results in predicting blood glucose peaks in healthy individuals.(3)To address the problem of long-term dependence of time series data and the correlation among the collected features,a blood glucose trend prediction model based on Bidirectional Long Short Term Memory Network(Bi-LSTM)and attention mechanism was proposed.The method combined Convolutional Neural Networks,Bi-LSTM,and attention mechanism to explore the connection between each feature data,and multiple time steps blood glucose trend prediction was realized.Compared with other model structures implemented in the fusion experiment,this model effectively improved the accuracy of blood glucose trend prediction.(4)To address the problem that some of the feature data are prone to errors in the collection process,a prediction model combining wavelet decomposition and Gated Recurrent Unit(GRU)was proposed to predict future blood glucose trends only using CGM time series data.The model showed the best prediction performance in predicting blood glucose trends at multiple time steps when compared with several common methods for time series prediction.In this paper,all three proposed methods for predicting blood glucose trends were experimentally validated and analyzed against each other.The results show that the proposed methods can predict the blood glucose trend relatively accurately for different characteristic data requirements,which can helpful to play a greater role in the fine management of blood glucose level.