Multi-Product Calibration Models of near Infrared Spectra of Foods

Abstract

Maintenance of multivariate calibration models can be laborious when using near infrared (NIR) for quantitative analysis, thus to save time on updates of the multivariate calibration models it is worthwhile to investigate if multi-product models for the content of a given constituent in many different foods can be developed. In this study eight different food products with known fat content (%) were analysed with NIR spectroscopy in the range from 850 to 1048 nm and the predictive performance of calibration models developed on each product was compared with multi-product calibration models including all eight food products. Different pre-processing techniques (multiplicative scatter correction, first and second order derivatives, extended inverted signal correction, standard normal variate transformation and second derivative combined with multiplicative scatter correction) and five linear and non-linear calibration models (partial least squares regression, neural networks and three local regression techniques) were evaluated. Also, simple single-spectral match was tested. In total, 4382 samples were analysed covering all eight products and a fat content range of 0.03 to 86.25%. Two thirds of the samples were used for model development and one third of the samples were used for independent test set validation. Not unexpectedly, the results showed that multi-product models generally result in less accurate and precise predictions than one-product models. Of the calibration methods tested, the local regression models gave the smallest range-relative root mean square error of prediction for multi-product models. Multi-product models also worked well using neural networks, whereas partial least squares regression could not handle the non-linear trend in data. Both neural network and local regression models were found to be suitable methods for substituting several one-product models with one multi-product model. The relative prediction errors for the best local regression model were in the range 2.2 to 4.4% for the eight products. Predictions based on using the fat content of the sample in the calibration set with the most similar spectrum (smallest Euclidian distance) to the test sample as the predicted value were inferior to the neural network and local regression methods, but worked surprisingly well.