The prevalence of "grass-fed" labeled food products on the market has increased in recent years, often commanding a premium price. To date, the majority of methods used for the authentication of grass-fed source products are driven by auditing and inspection of farm records. As such, the ability to verify grass-fed source claims to ensure consumer confidence will be important in the future. Mid-infrared (MIR) spectroscopy is widely used in the dairy industry as a rapid method for the routine monitoring of individual herd milk composition and quality. Further harnessing the data from individual spectra offers a promising and readily implementable strategy to authenticate the milk source at both farm and processor levels. Herein, a comprehensive comparison of the robustness, specificity, and accuracy of 11 machine-learning statistical analysis methods were tested for the discrimination of grass-fed versus nongrass-fed milks based on the MIR spectra of 4,320 milk samples collected from cows on pasture or indoor total mixed ration-based feeding systems over a 3-yr period. Linear discriminant analysis and partial least squares discriminant analysis (PLS-DA) were demonstrated to offer the greatest level of accuracy for the prediction of cow diet from MIR spectra. Parsimonious strategies for the selection of the most discriminating wavelengths within the spectra are also highlighted.

Numerous statistical machine learning methods suitable for application to highly correlated features, as those that exist for spectral data, could potentially improve prediction performance over the commonly used partial least squares approach. Milk samples from 622 individual cows with known detailed protein composition and technological trait data accompanied by mid-infrared spectra were available to assess the predictive ability of different regression and classification algorithms. The regression-based approaches were partial least squares regression (PLSR), ridge regression (RR), least absolute shrinkage and selection operator (LASSO), elastic net, principal component regression, projection pursuit regression, spike and slab regression, random forests, boosting decision trees, neural networks (NN), and a post-hoc approach of model averaging (MA). Several classification methods (i.e., partial least squares discriminant analysis (PLSDA), random forests, boosting decision trees, and support vector machines (SVM)) were also used after stratifying the traits of interest into categories. In the regression analyses, MA was the best prediction method for 6 of the 14 traits investigated [curd firmness at 60 min, α S1 -casein (CN), α S2 -CN, κ-CN, α-lactalbumin, and β-lactoglobulin B], whereas NN and RR were the best algorithms for 3 traits each (rennet coagulation time, curd-firming time, and heat stability, and curd firmness at 30 min, β-CN, and β-lactoglobulin A, respectively), PLSR was best for pH, and LASSO was best for CN micelle size. When traits were divided into 2 classes, SVM had the greatest accuracy for the majority of the traits investigated. Although the well-established PLSR-based method performed competitively, the application of statistical machine learning methods for regression analyses reduced the root mean square error compared with PLSR from between 0.18% (κ-CN) to 3.67% (heat stability). The use of modern statistical machine learning methods for trait prediction from mid-infrared spectroscopy may improve the prediction accuracy for some traits.

No abstract