Optimization of discriminant partial least squares regression models for the detection of animal by-product meals in compound feedingstuffs by near-infrared spectroscopy.

This paper evaluates two multivariate strategies for classifying near-infrared (NIR) spectroscopic data for the detection of animal by-product meals (henceforth generically termed AbP) as an ingredient in compound feedingstuffs. Classification models were developed to discriminate between the presence and absence of animal-origin meals in compound feeds using two forms of discriminant partial least squares (PLS) regression: the algorithms PLS1 and PLS2. The training set comprised 433 commercial feeds, of which 148 contained AbP and the other 285 were stated to be AbP-free. Since the initial set contained unequal numbers of each class, the effect of this imbalance was analyzed by applying the same algorithms to a training set containing equal numbers of AbP-free and AbP-containing samples. The best classification model (97.42% of samples correctly classified), obtained with PLS2, that showed less sensitivity to the use of class-unbalanced sets, was externally validated using a set of 18 samples (10 AbP-containing and 8 AbP-free); all samples were correctly classified, except for one AbP-free sample that was classified as containing AbP (false positive). The results suggest that the application of PLS discriminant analysis to NIR spectroscopic data enables detection of AbP, a feed ingredient banned since the bovine spongiform encephalopathy (BSE) crisis; this confirms the value of NIRS qualitative analysis for product authentication purposes.