Data analysis strategies for quantitative near-infrared spectroscopy: Application to the measurement of organic solvents in water and the determination of glucose in biological matrices

Near-infrared (near-IR) spectroscopy has been widely used for the determination of chemical species in environmental and biological applications. Because of the relatively low absorbance of analyte spectral bands and interferences from overlapping absorption bands in the near-IR region, the choice of data analysis strategies is critical to the successful implementation of an analysis based on near-IR spectroscopy. In this dissertation, data analysis strategies are developed for use in two applications: the quantitative monitoring of aqueous samples for organic contaminants and the determination of glucose in biological matrices.; The feasibility for the development of an on-line screening technique that will allow quantitative determination of organic solvents in water over the concentration range of 1-100 ppm is demonstrated. A genetic algorithm (GA)-based procedure is employed to optimize the calibration model parameters based on a combination of digital filtering and partial least-squares (PLS) regression.; The research on the determination of glucose in biological matrices is part of the work in our laboratory to develop data analysis methods for the in vivo, noninvasive measurement of blood glucose. Nonlinear model building strategies such as quadratic PLS (QPLS), stepwise QPLS, and artificial neural networks (ANN) are investigated and compared to linear PLS regression. These nonlinear models are observed to perform better than linear PLS models for the determination of glucose in an aqueous matrix of bovine serum albumin and triacetin. An enhanced GA-based wavelength selection procedure with random selection of a small number of initial wavelengths is developed for the joint optimization of the wavelengths used and the number of PLS factors employed in building an optimal calibration model. This GA-based procedure is observed to significantly reduced the number of wavelengths selected in the optimal models. Data processing methods are also developed for the direct analysis of single-beam spectra. The use of single-beam spectra has the advantage of no requirement for the collection of background spectra for use in processing spectra of samples whose concentration is sought. With the removal of intensity variations in the single-beam spectra through the use of preprocessing methods (i.e., log transform, multiplicative signal correction (MSC), normalization), the direct analysis of single-beam spectra is demonstrated as a viable approach for the determination of glucose in biological matrices.