Research And Application Of Dynamic Blood Glucose Prediction Based On Active Learning

Diabetes is a chronic metabolic disease,characterized by high blood glucose,which has many patients and cannot be cured.It has been listed as the third largest disease after cardiovascular disease and tumor.blood glucose is the only standard for diagnosis and tracking of diabetes.Currently,artificial intelligence algorithms such as machine learning and deep learning,combined with traditional sensor based data feature quantity modeling methods,are gradually applied to blood glucose measurement.However,due to the limited medical data resources in the medical field,data driven algorithm models led by deep learning are not suitable for high-precision prediction of blood glucose values,and relying solely on sensor features for real-time blood glucose prediction is not reliable.Therefore,this thesis analyzes and studies the real-time prediction and dynamic time series prediction of human blood glucose under small samples,and proposes a dynamic blood glucose prediction method based on active learning.This article mainly does the following three aspects of work.(1)This thesis proposes an active learning algorithm and blood glucose prediction model based on Bhattacharyya distance measurement.There are two existing methods for predicting blood glucose.One is a traditional mathematical calculation method based on formula theory and experience,but it cannot maintain high accuracy in the face of complex human body conditions;The other is a method based on machine learning and deep learning,which uses a large amount of data for model training to achieve blood glucose prediction.However,it relies heavily on data and performs poorly under the existing medical experimental conditions with insufficient data.This article utilizes the concept of active learning,which can actively select samples with higher information content for training.It can fully utilize sensor feature data,combined with Bhattacharyya distance,to achieve a balance between uncertainty and representativeness,and jointly apply it to blood glucose prediction.In the selection of different blood glucose prediction models,this article compares other different algorithm models and traditional prediction models.The experimental results show that the active learning algorithm model based on Bhattacharyya distance metric outperforms most CGM prediction methods in accuracy.(2)This thesis proposes a blood glucose trend prediction method based on the GRU time series prediction model.The multivariate data used in this article contain time-related components,which have a chronological relationship.In the real-time blood glucose monitoring environment,previous status also affects current and future changes in blood glucose status.Based on the characteristics of multivariate data,this thesis introduces a feature quantity related to time series to supplement the overall data information.First,calculate the user’s blood glucose value at the current time point through an active learning prediction algorithm,and then use the GRU time series prediction model to train and find temporal correlations from the current historical data.Finally,use the model to predict and judge the possible changes in the future blood glucose value,in order to achieve a carpenter’s blood glucose trend.The experimental results show that time series features and time series models can fully mine the effective information of sensor data and achieve a more accurate description of blood glucose prediction.(3)Based on active learning and temporal model prediction methods,this thesis designs and implements a CGM real-time blood glucose prediction system.The modular design is the structural basis of the system.Users can have a more specific understanding of their own blood glucose situation through switching models,and can also view historical blood glucose values to better coordinate medical diagnosis and treatment work.The system test results show that the system has the characteristics of practicality and reliability,and verify the effectiveness and practicality of this method from an engineering perspective.