Abstract
The effects of spectral resolution on the performance of partial least-squares (PLS) calibration models were studied in the context of an analysis for physiological levels of glucose by Fourier transform near infrared (NIR) spectrometry in the combination region of 5000 to 4000 cm−1. Employing a simulated biological matrix of varying levels of triacetin and bovine serum albumin in phosphate buffer, two spectral data sets were evaluated. Single-sided interferograms consisting of 16,384 points collected over a 15,801 cm−1 bandwidth were Fourier processed to yield spectra with a nominal 2 cm−1 point spacing. These interferograms were then progressively truncated before Fourier processing to form sets of spectra with reduced resolution. The lower-resolution spectra were also computed with varying levels of zero-filling to evaluate the effects of spectral dimensionality on the PLS models. Across the levels of resolution and zero-filling, 21 parallel groups of spectra were computed with each of the two data sets. Calibration studies were then performed with spectra in absorbance units. Predictive models for glucose were optimised in terms of spectral range, the number of PLS latent variables and the parameters of a pre-processing Fourier filter. Models were evaluated for their performance in both short-term and long-term predictions. Through this research, it was found that models based on spectra with a nominal resolution of 16 cm−1 performed as well as those computed with higher-resolution spectra. Over the range of 1–20 mM glucose, these models achieved values of the standard error of prediction (SEP) in the range of 0.5 mM for short-term predictions and 0.8 mM for predictions with spectra collected more than 200 days after the calibration data. These results are significant because they help to define the required instrumental characteristics of a potential dedicated NIR glucose sensor.
Get full access to this article
View all access options for this article.