Research On Data Mining Of Blood Glucose Spectrum Based On Machine Learning

Data mining technology based on machine learning is developing rapidly and playing an important role in the Internet and data analysis.It can dig out effective information from large and noisy data to solve problems.As one of the important indicators to evaluate human health,the high-precision non-invasive detection of blood glucose is of great significance to discovery and control diseases,and it is a hot issue that attracts much attention in the academic and industrial.Due to the low content of blood glucose in human and the spectral signal of blood glucose is too weak to extract,the prediction effect of noninvasive blood glucose concentration is not good,so how to achieve high-precision noninvasive blood glucose detection has become a typical "bottleneck" problem.In this thesis,near-infrared(NIR)spectral signals are taken as the research object,and machine learning methods such as support vector machine and deep learning are used to carry out related studies on the concentration prediction of non-invasive glucose spectral signals.The main research contents are as follows.1)A non-invasive blood glucose concentration prediction algorithm based on ε-support vector regression machine(ε-SVR)is designed.A discrimination algorithm is proposed to select wavelengths,and the effects of different sample sizes and feature dimensions on the ε-SVR model are investigated.Compared with the traditional partial least squares regression prediction model,the experimental results show that the ε-SVR model is effective and feasible in the prediction of NIR non-invasive blood glucose concentration.2)A non-invasive blood glucose concentration prediction algorithm based on deep learning is designed.Firstly,the non-invasive blood glucose concentration prediction is studied at seven specific wavelengths.The deep belief network(DBN)is used to extract the spectral characteristics of blood glucose,and the DBNSVR prediction model is established.Then,the effects of different sample sizes and spectral feature dimensions(i.e.,DBN network structure)on DBNSVR model accuracy are explored.Finally,the prediction accuracy of ε-SVR model and DBNSVR model are compared and analyzed.The results show that when using DBN network,the root mean square error of the prediction concentration of Volunteer 1 and Volunteer 2 decreased by 72.33% and 76.06%,the correlation coefficient increased by 13.99% and 39.81%,and the Clark error grid analysis results improved by 6.28% and 7.2%,respectively.The results show that this algorithm plays a promoting role in non-invasive blood glucose concentration prediction and it can effectively improve the prediction accuracy of the DBNSVR model.In addition,the algorithm has a good performance on the independent data sets of two volunteers,which further shows the generalization ability of DBNSVR model.