Calibration transfer in Fourier transform infrared remote sensing and in the near-infrared spectroscopic determination of clinical analytes

Of the many instrumental technologies provided by analytical chemistry, the chemical selectivity, sensitivity, and flexibility provided by Fourier transform infrared (FT-IR) spectrometry makes it a common choice in broad classes of applications requiring a fast, non-destructive sensing technique. Arising as a practical technology after the arrival of computers, the analysis of the complex digitized chemical information provided by FT-IR spectrometry remains intimately intertwined with computer-based techniques, beginning with the Fast Fourier Transform itself. Like many other modern analytical methods, FT-IR spectrometry is capable of producing tremendous quantities of data that often require complex analysis techniques. These techniques attempt to leverage data into information that can be used to make practical decisions. Techniques discussed in this work in the computer-based analysis of FT-IR data include digital signal processing, digital filtering, pattern recognition, multivariate calibration, and optimization. Whether in the automated qualitative determination of air pollutants in industrial stack monitoring, or a highly sensitive quantitative determination of clinical analytes, when applied to FT-IR data, these data analysis techniques seek the suppression of interfering signals and the extraction of analyte information, allowing these important applications to be realized.; Techniques applied to FT-IR data are becoming more powerful in extracting the maximum amount of chemical information from the instrumental signal. Often, however, signals not related to the analyte such as background features, temperature information, and spectrometer-specific information become included in the multivariate models used in qualitative or quantitative prediction. These interferences can represent orders of magnitude more variation in the FT-IR data than the changes in the signal arising from the analytes of interest. If the model includes any of these kinds of information, the resulting prediction can be of poor quality if any aspect of the analytical measurement environment has changed. The large amount of work required in collecting and analyzing training data used to build these models then needs to be repeated for the different experimental conditions, environment, or different spectrometer.; The work presented in this dissertation focuses on exploring these calibration transfer issues while overcoming difficult challenges posed by the analysis of FT-IR data from two different applications. First, qualitative passive FT-IR remote sensing of several analytes is performed with field and laboratory data collected with different spectrometers. The role of digital filtering is examined in removing instrument-specific signals and allowing data from either instrument to be predicted with a single model. Second, quantitative near-infrared analysis of clinical glucose and urea samples is performed while examining the performance of models with data collected over two years apart, as well as with data collected with two different spectrometers. Methodologies from enhanced digital filtering to the creation of more robust models are described for improving the calibration transfer results in each of these important projects.