RAMAN AND NIR SPECTROSCOPIC METHODS FOR DETERMINATION OF TOTAL DIETARY FIBER IN CEREAL FOODS - UTILIZING MODEL DIFFERENCES

This work evaluates the complementarity in the predictive ability of t hree Raman and three near-infrared reflectance (NIRR) partial least-sq uares regression (PLSR) models for total dietary fiber (TDF) determina tions of a diverse set of ground cereal food products. For each spectr al type (R or N), models had previously been developed from smoothed ( D0), first-derivative (D1), or second-derivative (D2) spectral data. T he NIRR and Raman models tend to have very different sets of outliers and uncorrelated errors in TDF determination. For a single spectral ty pe, the prediction errors of various preprocessing methods are partial ly complementary. The samples are very diverse in terms of composition , but the main problem groups were narrowed to high-fat, high-bran, an d high-germ samples, as wed as and those containing synthetic fiber ad ditives. Raman models perform better on the high-fat samples, while NI RR models perform better with high-bran and high-synthetic samples. Ra man models were better able to accommodate a wheat germ sample, even t hough this sample type was poorly represented by the calibration set. Two methods are presented for utilizing the complementarity of the spe ctral and processing techniques: one involves simple averaging of pred ictions and the other involves avoidance of outliers by using statisti cs generated from the sample spectrum to choose the best model(s) for determination of the TDF value. The single best model (N-D1) has a roo t-mean-squared prediction error of 2.4% TDF. The best model of predict ion averages yields an error of 1.9% (combining N-D0, N-D1, N-D2, R-D0 , and R-D1). An error of 1.9% was also obtained by choosing a single p rediction from the six models by using statistics to avoid outliers. W ith the selection of the best three models and averaging their predict ions, an error of 1.5% was achieved.