RAMAN AND NIR SPECTROSCOPIC METHODS FOR DETERMINATION OF TOTAL DIETARY FIBER IN CEREAL FOODS: A COMPARATIVE STUDY

Authors: Archibald, Douglas; Kays, Sandra; Himmelsbach, David; Barton Ii, Franklin

Submitted to: Journal of Applied Spectroscopy
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 7/5/1997
Publication Date: N/A
Citation: N/A

Interpretive Summary: This work studies the use of multiwavelength light reflection to determine the dietary fiber content of a wide range of cereal food products. It involves measuring light reflection from a representative set of cereal food samples with known dietary fiber content and subsequently developing a mathematical calibration model to allow prediction of dietary fiber from the optical information. The main goal is to substitute a simple and rapid optical method for the wasteful and time-consuming wet chemical method of dietary fiber determination that is currently in use by regulatory and commercial laboratories. Companion goals are to understand the chemical nature of dietary fiber for nutritional research and to understand the interaction of light with agricultural products so that other applications of the optical methods may be facilitated. For these purposes, two types of light interaction were studied: near-infrared-reflectance and Raman scattering. The study also reports optimization of a variety of mathematical techniques for extracting the dietary fiber information from the optical data. The main novel finding is that it is possible to optically determine dietary fiber for a wide range of products using Raman scattering. The measurement accuracy is better than the information found on the product labels used in the study.

Technical Abstract: Partial least squares regression (PLSR) was used to generate three Raman and three near-infrared reflectance (NIRR) models for the spectroscopic determination of total dietary fiber (TDF) of a wide variety of cereal foods. To allow comparison of the spectral techniques, both analyses used the same sets of samples (ncal=63, nval=63). Six models were optimized by full leave-one-out cross-validation (CV), including a smoothed, first and second derivative model for each spectral technique. Both kinds of raw spectral data required correction of interfering baseline and amplitude variations. Derivative preprocessing generally reduced the number of latent variables (LVs) for both spectral types, and significantly reduced the CV error of the Raman models. Raman models required 6 to 9 latent variables (LVs) while NIRR models required 10 to 14 LVs. The root-mean-squared CV model errors were 2 - 2.3% TDF for all six models, and dthe three Raman models had root-mean-squared prediction errors (RMSEP) in the range 2.8 - 3.2% TDF, with the best model being generated from second derivative data. First derivative data provided the best NIRR-model and for all three NIRR models the RMSEP spanned 2.4 - 2.9%. For some types of samples, it is suggested that the Raman method is limited by its sampling technique and could be improved with more densely packed, larger-area specimens. The regression vectors of the Raman models seem more easily interpretable than NIRR models. Either spectral method appears capable of achieving an acceptable level of error; TDF reference method precision was 0.68% TDF, while the product label information had an error of 2.8% TDF relative to the reference.