Variable selection for quantitative determination of glucose concentration with near-infrared spectroscopy

Near-infrared spectroscopy has been used in combination with multivariate calibration techniques such as partial-least squares regression (PLSR) to quantify glucose concentration in various media. However, for reasonable prediction capability in measuring glucose many calibration samples are needed. In addition, spectroscopic data often contain over 1000 data points, presenting a very large data matrix for calibration. It is desirable to reduce the available data to contain only the information necessary for accurate prediction of chemical concentration before PLSR is applied. This will eliminate noisy variables and consequently the data can be processed more quickly and efficiently. A variable selection method that reduces prediction bias in single factor partial least squares regression models was developed and applied to near-infrared absorbance spectra of glucose in two different media: pH buffer and cell culture medium. Comparisons between calibration and prediction capability for full spectra and reduced sets were completed, resulting in statistically equivalent mean squared errors. The number of response variables needed to fit the calibration data and accurately predict concentrations from new spectra was reduced in each case. The algorithm correctly chose the glucose peak areas as the informative variables and computation time was decreased by an order of magnitude.