Development of Robust Calibration Models Using Support Vector Machines for Spectroscopic Monitoring of Blood Glucose

摘要

Sample-to-sample variability has proven to be a majornchallenge in achieving calibration transfer in quantitativenbiological Raman spectroscopy. Multiple morphologicalnand optical parameters, such as tissue absorption andnscattering, physiological glucose dynamics and skin heterogeneity,nvary significantly in a human populationnintroducing nonanalyte specific features into the calibrationnmodel. In this paper, we show that fluctuations ofnsuch parameters in human subjects introduce curvedn(nonlinear) effects in the relationship between the concentrationsnof the analyte of interest and the mixturenRaman spectra. To account for these curved effects, wenpropose the use of support vector machines (SVM) as annonlinear regression method over conventional linearnregression techniques such as partial least-squares (PLS).nUsing transcutaneous blood glucose detection as annexample, we demonstrate that application of SVM enablesna significant improvement (at least 30%) in cross-validationnaccuracy over PLS when measurements from multiplenhuman volunteers are employed in the calibrationnset. Furthermore, using physical tissue models withnrandomized analyte concentrations and varying turbidities,nwe show that the fluctuations in turbidity alonencauses curved effects which can only be adequatelynmodeled using nonlinear regression techniques. Thenenhanced levels of accuracy obtained with the SVM basedncalibration models opens up avenues for prospectivenprediction in humans and thus for clinical translation ofnthe technology.