Short communication: Potential prediction of vitamin B₁₂ concentration based on mid-infrared spectral data using...

The purpose of this study was (1) to predict the quantitative concentration of vitamin B₁₂ in milk using mid-infrared (MIR) spectrometry, and (2) to evaluate the potential of MIR spectra to discriminate different clusters of records based on their B₁₂ concentration. Milk samples were collected from 4,340 Holstein cows between 3 and 592 d in milk and located in 100 herds. Samples were taken using in-line milk meters and divided into 2 aliquots: one for MIR spectrometry and the other for B₁₂ concentration reference analyses by radioassay. Analyses were performed on 311 selected spectral wavelengths. A partial least squares regression model was built to quantify B₁₂ concentration. Discriminant analysis was used to isolate B₁₂ concentration clusters. A B₁₂ concentration threshold was set at 442 ng/dL, because this represents the cutoff value for a 250-mL glass of milk to fulfill 46% of the daily vitamin B₁₂ recommended dietary allowance for individuals 14 yr or older. For each analysis, records coming from two-thirds of herds were used to calibrate prediction equations, and the remaining records (one-third of herds for validation) were used to assess the prediction performance. In the case of discriminant analysis, validation sets were divided into evaluation sets (one-third of herds) to obtain alternate probability cutoffs and in test sets (two-thirds of herds) to validate equations. Spectral and B₁₂ concentration outliers were identified by calculating standardized Mahalanobis distance and with a residual analysis, respectively (n = 3,154). Regarding quantitative B₁₂ concentration, cross-validation and validation coefficients of determination averaged 0.51 and 0.46, respectively, which are relatively low, which would limit the potential use of the developed quantitative equations. In addition, root mean square errors of prediction of cross validation and validation sets averaged 88.9 and 94.7 ng/dL, respectively. Area under the receiver operating characteristic curve of test sets averaged 0.81 based on the 442 ng/dL threshold, which could be considered to represent good accuracy of classification. However, the false discovery rate averaged 36%. In summary, models predicting quantitative B₁₂ concentration had low cross-validation and validation coefficients of determination, limiting their use, but the proposed discriminant models could be used to identify milk samples with naturally high B12.