Research On Prediction Algorithm Of Non-invasive Blood Glucose Concentration Based On Deep Learning

Recently,deep learning is an essential theoretical development of machine learning.It is mainly used to learn deep characteristics and the inherent laws of sample data.Intelligent algorithms based on deep learning have been brilliant in many fields such as image search and recognition technology,natural language processing,biomedicine and so on.Blood glucose is an important indicator of human health.At present,usually use invasive methods to monitor clinical blood glucose concentration.Non-invasive blood glucose concentration measurement technology has been developed,It is the expectation of many patients.Therefore,how to achieve high-precision non-invasive blood glucose concentration detection has become a hot issue of concern for relevant industries and experts and scholars.In recent years,the nondestructive testing technology based on near-infrared spectroscopy has developed rapidly,and it has blossomed in industries,life sciences,agriculture,clinical medicine and other application fields.In this paper,non-invasive blood glucose concentration prediction research is carried out by acquiring non-invasive blood near-infrared spectral data and extracting non-invasive blood glucose concentration characteristics based on deep learning method.The main research contents are as follows:1)Considering the problem of low prediction accuracy of non training data in the same individual and concentration,we consider deep mining the features of non-invasive blood glucose concentration.Therefore,in order to deeply mine the characteristics of non-invasive blood glucose concentration,the label sensitivity feature wavelength algorithm(LS algorithm)and the deep confidence network feature extraction algorithm(DBN algorithm)were proposed,and the LSDBNSVR blood glucose concentration prediction model was constructed.The effects of the number of nodes in the DBN hidden layer and the feature dimension on the prediction performance of the model were compared and analyzed.The results showed that the correlation coefficient between the predicted value of the LSDBNSVR algorithm and the true concentration value was as high as 99.9%,The mean square error(MSE)is less than 0.001,which indicates that the depth feature obtained based on LS algorithm and DBN algorithm plays a positive role in improving the accuracy of noninvasive blood glucose concentration prediction.2)The problem of low concentration prediction accuracy and poor generalization ability for the same individual who has never participated in training.In order to further improve the generalization ability of the in-depth learning model for non-invasive blood glucose concentration prediction,a feature extraction model based on the deep twin network and enhanced measurement loss(RMLoss)was proposed,and a non-invasive blood glucose concentration prediction model was established based on SVR.The influence of the selection of network structure,the setting of output dimension and the selection of hyper-parameters on the prediction results was explored,The prediction performance is the best when K=0.5.The mean square error of the non-invasive blood glucose concentration prediction model based on the deep twin network and SVR proposed in this paper is reduced to 0.024,which is 49% higher than the SVR correlation coefficient.It provides a solution strategy based on the deep learning algorithm for the problems that need to be solved.