Improving the robustness of multivariate calibration models for the determination of glucose by near-infrared spectroscopy

by Kramer, Kirsten Elizabeth.

Near-infrared spectroscopy has proven to be one of the most promising techniques for the development of a noninvasive blood glucose monitoring system for diabetic patients. In this work, Fourier transform infrared (FT-IR) transmission measurements of the combination band region (4000 – 5000 cm-1) were analyzed for samples containing glucose (analyte) in a matrix of bovine serum albumin and triacetin (models for proteins and fats), all spanning physiological levels relevant for a diabetic patient. The first part of the study investigated the required spectral point-spacing for accurate detection of glucose. This was studied by systematically truncating interferograms before Fourier transforming them to single-beam spectra. A set of training data (70 samples) was collected for multivariate calibration using partial least-squares (PLS) and an external prediction set was used to verify the success of modeling glucose quantitatively. It was found that a relatively large point-spacing (16 cm -1) was successful for prediction of glucose, meaning that a shorter interferogram could be collected. The second part of the study involved collecting interferograms such that the spectral resolution was 16 cm-1, and investigating methods to extend the usefulness of calibration models for long-term data collection. Near-infrared spectroscopy often suffers from weak signals that are overwhelmed by significant instrumental drift, meaning that calibration models tend to be unsuccessful for data collected several days or months outside the calibration. For updating the calibration models, a set of 50 backgrounds containing only matrix constituents without analyte was collected on each analysis day, and used to 2 update the original calibration model so that instrumental drift features were incorporated into the model. Background updating was found to be successful in single-beam format, producing a background-augmented (BA) PLS model that significantly improved single-beam data analysis. The standard error of prediction using the original model (PLS) and the updated model (BA-PLS) were 13.4 and 0.79 mM glucose, respectively, for a prediction set taken 176 days outside of the calibration. The matrix data also allowed for studies in background selection methods for absorbance computations as well as adaptive digital filtering that was guided by the background data. Abstract Approved : ______________________________________ Thesis Supervisor ______________________________________ Title and Department ______________________________________ Date