Extraction Of Effective Signal In Noninvasive Blood Glucose Sensing With Nearinfrared Spectroscopy

Relative measurement of near-infrared(NIR)spectroscopy is of great significance for the high precision detection of human blood glucose concentration in vivo.The differential floating reference measurement is a universal method.However,it is difficult to determine the radial detection positions and extract the effective information from differential signals in practical use.To solve these problems,differential floating reference based on NAS(Net Analyte Signal)-VIP(Variable Importance in Projection)-SPXY(Sample set Partitioning based on joint X-Y distances)-PLS(Partial Least Square)method was proposed and the feasibility was investigated by Monte Carlo simulation,in vitro and in vivo experiments.Firstly,the light propagation in three-layer-skin model were simulated by the Monte Carlo program.The inside and outside distances of differential detection fiber from the central light source were determined,which were 0.6 mm and 2.5 mm,respectively,according to the distribution rules of glucose and interference(e.g.light source drift)with radial distance.The inner ring can be used to measure useful signal with higher sensitivity to glucose change,while the outer one can be used as a reference to build the noise subspace to extract the NAS.In addition,VIP and SPXY methods were applied for the optimization of wavelength and calibration set,respectively.Then the PLS model was established between the glucose and NAS.Results showed that,the wavelength dimension was dropped 54.09%,and RMSEC and RMSEP of the model were reduced by 18.52% and 20.42%,respectively,and the correlation coefficient was also improved after these optimizations.Secondly,in vitro experiment was conducted with 3%Intralipid solution.The noise subspace was formed by the spectrum of outer ring to extract the glucose of NAS.The influence of different light drift levels on the NAS was investigated and the NAS-VIPSPXY-PLS model of glucose was established.It was shown that,the wavelength dimension was dropped 60.98%,and the RMSEC and RMSEP of the specified model were reduced by 39.37% and 42.25%,respectively.Finally,in vivo experiment with six healthy volunteers was carried out.The NASVIP-SPXY-PLS model of blood glucose concentration was established.the wavelength dimension was dropped more than 42% after optimization.Moreover,the prediction error of individual models had reduced at different degrees with an average RMSEP of 0.545 mmol/L.And Clarke Error Analysis showed that the predictions were fallen in the region A.