Regression Analysis Of Near-infrared Non-Invasive Blood Glucose Concentration Based On Deep Learning

In recent years,deep learning has achieved obvious and successful applications in image processing,speech recognition,etc.,and has promoted the development of intelligent technology research based on deep learning in many fields.The weak spectral signal,as well as interference factors such as changes in human physiological conditions and different data collection sites,lead to high-precision non-invasive blood glucose concentration detection based on spectral technology has always been a typical "stuck neck" problem.The rapid development of deep learning technology brings hope to solve these problems.This paper uses deep belief network,deep siamese residual network,and support vector machine to predict non-invasive blood glucose concentration.The main research contents are as follows.(1)Non-invasive blood glucose concentration regression analysis based on deep belief network.The deep belief network designed in this paper is composed of 3-layer boltzmann machine stacking,which realizes the extraction of original spectral data to deep features,and further establishes support vector machine regression.Compared with the prediction results of blood glucose concentration without using deep features,the average absolute errors of the three designed experimental results were reduced by 19%,20%,and 10%,respectively.The root mean square error is reduced by 48%,35%,and 20%,respectively.(2)Non-invasive blood glucose concentration regression analysis based on the deep siamese residual network.The deep siamese residual network is composed of two residual networks of the same structure with shared parameters.After calculation such as convolution pooling of the residual network,the original data is converted is a32-dimensional deep siamese feature,and the concentration prediction is further performed by the support vector machine model.In the three groups of experimental analysis of the effect of different data collection sites on concentration regression,compared with the original features,the average absolute error of support vector machine concentration prediction based on the deep twin feature was reduced by 9%,11%,and 29%,respectively,and the root mean square error were reduced by 16%,26%,and 43%,respectively.In the three groups of experimental analysis of the effect of different body states on the concentration regression,compared with the original characteristics,the average absolute error of support vector machine concentration prediction based on deep siamese features is reduced by 15%,22%,and 20%,respectively,and the root mean square error is reduced by25%,23%,and 29%,respectively.