Research On Blood Glucose Estimation Based On Signal Processing Technology

Diabetes mellitus has become a public health problem that many countries in the world have to pay attention to.Because of its long illness and many complications,which consumes a lot of public health resources every year.The daily monitoring of blood sugar level is an important way for diabetic patients to assist treatment and reduce or delay complication.During to the reason of invasive blood glucose detection technology that it can not achieve continuous detection and ignore the peak blood glucose problems,non-invasive blood glucose detection technology has become the focus of many research institutions.Inspired by the research of the predecessors,the smart bracelet developed by Glutrac Information Technology(Shenzhen)Co.,Ltd.independently was used to collect PPG and ECG signals of five volunteers at one hour before meals and one hour after meals for two consecutive weeks.The other five indies such as blood diastolic pressure,blood systolic pressure,heart rate,respiratory rate and oxygen value are collected by other instruments.At the same time,their reference blood glucose was obtained by invasive methods.In order to improve the signal preprocessing method,empirical model decomposition method is employed to remove the baseline of the collected PPG signal and ECG signal at first.And then the singular spectrum analysis algorithm is used to denoise the signal.Finally,the feature of PPG and ECG signal are extracted separately.At the end of the paper,four machine learning algorithm,elastic network,Adaboost,XGBoost and LightGBM,are employed to model individual and multi-person with signal model.Moreover,the integration algorithm is used to combine four algorithms for further improvement.Comparative experiments in this paper shows that,the accuracy of individual model is better than that of multi-person model.The accuracy of the integrated model is better than that of signal model because it combines the prediction result of multiple model.Whether the integration model based on personal modeling or the integration based on multi-person modeling,the percentage of point in area A of Clark error grid can reach more than 80%.