Simulation And Experiment Study On Outliers Determination In The Spectral Measurement By Near-infrared Diffuse Reflectance

The complexity of diffuse reflectance spectroscopy frequently makes outliers which cannot correctly reflect the glucose specific information in near-infrared non-invasive blood glucose monitoring.Thus,it is necessary to determine and eliminate outliers for establishing a high quality model to predict blood glucose concentration.In this thesis,the spectral signals including the information of glucose concentrations,temperature change and drift were simulated by Monte Carlo with the three-layer-skin model firstly.The distribution and wavelength characteristics of the spectral signals were obtained,guiding the selection of the source-detector separation and wavelength.Then,I discussed the types of outliers appeared in the modeling process and the corresponding indicators.After that,a 3D coordinate method was proposed based on chemical value residual,mahalanobis distance and spectral residual,and used to determine outliers in the Monte Carlo simulation and experiment data.Firstly,for the three-layer-skin model,the Monte Carlo simulation was employed to analyze the distribution and wavelength characteristics of the spectral signals including the information of glucose concentrations,temperature change and drift on the source-detector separations.The simulation results showed that the preferred source-detector separations were 0.8~1mm and near-infrared wave bands were1000~1400nm or 1500nm~1700nm in the near-infrared blood glucose monitoring.Secondly,for the outliers frequently appearing in the modeling process of near-infrared blood glucose monitoring,a 3D coordinate method was proposed based on chemical value residual,mahalanobis distance and spectral residual.The simulation data including abnormal chemical value and spectra was obtained based on Monte Carlo by setting the abnormal situations such as human error,extreme compositions in the sample and temperature changes,and the 3D coordinate method,Hotelling T~2 and MCCV were applied to determine and eliminate the outliers.Results showed that,after eliminating outliers based on the three methods,RMSEC of the models was reduced by95%,88%and 88%,respectively.Finally,in vivo experiments were conducted with three healthy volunteers,and three sets of experiment data were obtained by simultaneously measuring the blood glucose concentration and diffuse reflectance spectra.The three methods were applied respectively to determine and eliminate the outliers.Results showed that,after eliminating outliers based on the 3D coordinate method,Hotelling T~2 and MCCV,the mean of RMSEC was reduced from 1.48mmol/L to 0.19mmol/L,1.02mmol/L and0.65mmol/L,decreased by 83%,51%and 13%,respectively.These results indicate that the proposed 3D coordinate method is more appropriate for determining outliers in near-infrared non-invasive blood glucose monitoring.