Spectral Simulation Protocol for Extending the Lifetime of Near-Infrared Multivariate Calibrations

Abstract

Multivariate calibration models based on synthetic single-beam near-infrared spectra are used to demonstrate the ability to maintain viable calibrations over extended time periods. Glucose is studied over the physiological concentration range of 1−30 mM in a buffered aqueous matrix containing varying levels of alanine, ascorbate, lactate, urea, and triacetin. By employing a set of 25 test samples measured 23 times over a period of 325 days, partial least-squares (PLS) calibrations based on synthetic spectra are observed to outperform conventional calibrations that use a set of 64 measured calibration samples. The key to the success of this approach is the use of a set of spectra of phosphate buffer collected on each prediction day to construct synthetic calibration spectra that are specific to that day. This allows the incorporation into the calibration model of nonanalyte spectral variance that is unique to a particular day. In this way, the adverse effects of instrumental drift or other sources of spectral variance on prediction performance can be minimized. Through the application of this methodology, values of the standard error of prediction (SEP) for glucose concentration are maintained to a range of 0.50−0.95 mM and an average of 0.68 mM over the 325 days of the experiment. These results are significantly better than those obtained with conventional models based on measured calibration samples. Over the same time period, a PLS model based on measured calibration spectra in absorbance units produced values of SEP that ranged from 0.41 to 2.02 mM and an average of 1.23 mM.