Strategies for coupling digital filtering with partial least-squares regression: Application to the determination of glucose in plasma by Fourier-transform near-infrared spectroscopy

Abstract

Protocols are established for coupling digital filtering techniques with partial least-squares (PLS) regression for use in constructing multivariate calibration models from Fourier transform near-infrared absorbance spectra. Calibration models are developed to predict glucose concentrations in bovine plasma samples. Employing a calibration data set of 300 spectra collected from 55 plasma samples and 3 plasma lots, individual calibration models are developed based on four spectral ranges selected from the region 5000-4000 cm-1. A separate test set of 69 spectra collected from 14 plasma samples is used to evaluate the computed models. Gaussian-shaped bandpass digital filters are implemented by use of Fourier filtering techniques and employed to preprocess spectra to remove variation due to the background absorbance of the plasma matrix. PLS regression is used with the filtered spectra to compute calibration models for glucose. The optimization of the filter bandpass parameters is explored through the use of response surface methods. Through these optimization studies, calibration models are developed that achieve standard errors of estimate and standard errors of prediction in the range 0.4-0.5 mM across the concentration range of 2.5-25.5 mM. It is determined that the use of digital filtering as a preprocessing step significantly improves the performance of the resulting calibration models, minimizes the importance of spectral range in the calibration model development, and reduces the required number of PLS factors in each model.