Search

By controlling the blood glucose levels of diabetics permanent diabetes-related problems such as blindness and loss of limbs can be delayed or even avoided. Therefore, many researchers have aimed at the development of a non-invasive sensor to monitor the blood glucose level continuously. As non-invasive measurements through the skin, the ear lobe or the gums have proven to be either unreliable or impractical, attention has recently shifted to minimally invasive sensors which measure the glucose content in serum or interstitial fluid. Thanks to the development of on-chip spectrometers minimally invasive, implantable devices are coming within reach. However, this technology does not allow to acquire a large number of wavelengths over a broad range. Therefore, the most informative combination of a limited number of variables should be selected. In this study, Interval PLS (iPLS), Variable Importance in Projection (VIP), Uninformative Variable Elimination (UVE), Bootstrap-PLS coefficients, Moving window, CorXyPLS, Interval Random Frog-PLS and combinations of these methods were used in order to address the question whether the short wave band (800–1500 nm), first overtone band (1500–1800 nm), the combination band (2050–2300 nm) or a combination of them is the most informative region for glucose measurements and which wavebands should be measured within these wavelength ranges. The three different data sets employed focus on the determination of (1) glucose in aqueous solutions over the 1–30 mM range in presence of urea and sodium D -lactate, (2) glucose in aqueous solutions over the 2–16 mM range in presence of icodextrin and urea and (3) glucose in human serum samples. The best results for the first, second and third data sets were obtained by selecting 40, 130 and 20 variables resulting in a PLS model with an RMSEP of 0.56, 0.59 and 1.5 mM, respectively. It was found that the first overtone band is most informative for aqueous solutions, while for glucose measurement of serum samples the combination band was found to be the better choice. [ABSTRACT FROM AUTHOR]

Near-infrared (NIR) spectra of human blood serum consist of overlapping strong absorption bands of water and serum proteins, which affect the ability of multivariate calibration models to predict glucose. Furthermore, serum proteins such as albumin and globulins undergo a glycation reaction by forming covalent bonds with freely available glucose molecules in the serum. In diabetic individuals with poor glucose control, more and more serum protein molecules react with glucose, resulting in a high glycated protein concentration. The glucose molecules covalently bonded to serum proteins might contribute to the overall glucose signal acquired by NIR spectroscopy. This might affect the prediction ability of multivariate calibration models such as partial least squares regression (PLSR). In this study, we investigated the effect of total protein concentration and the glycated protein concentration in blood serum on the prediction ability of PLSR calibration models. Serum samples were subjected to ultra-filtration, and the PLSR model was built using NIR spectra of filtered serum solutions. Prediction performance was found to improve by 39-42% in absence of serum protein molecules. Various experimental data set designs were generated by carefully varying the glycated serum protein concentration in calibration and test sets of PLSR models. This investigation revealed that the impact of varying glycated protein concentration on the root mean square error of prediction was not drastic. To test the statistical significance of the prediction results, a multiple linear regression model was built. The glycated serum protein concentration was found to be statistically insignificant (p= 0.86) in predicting glucose concentration. Overall, it was concluded that the glycated serum proteins do not affect the glucose prediction accuracy of PLSR models using NIR spectra of human serum. [ABSTRACT FROM AUTHOR]