[en] The development of mid-infrared equations to predict the milk fatty acid (FA) content of milk allows prompt analysis of large numbers of samples. The first aim was to improve these predictions by comparing 6 statistical approaches. The second one was to validate the new equations using an independent sample set. The calibration set contained 239 spectrally different Belgian milk samples collected for over 2 years from several cows and breeds. FA were quantified by gas chromatography (GC). Statistical approaches tested were 1) partial least squares regression (PLS), 2) PLS and first derivative, 3) PLS and repeatability file (RF), 4) PLS, first derivative and RF, 5) PLS, second derivative, and 6) PLS, second derivative and RF. This last file contained spectra obtained from the same samples using 5 spectrometers. Cross-validation (CV) used 20 groups from the calibration set. Methods were compared using the ratio of the standard deviation of GC values to the standard error of CV (RPD). An external validation permitted a second comparison and was done using 362 samples collected for one year from multiple breeds and cows in Belgium, Ireland, and Scotland. Different RPD values were obtained by the 6 methods. Generally the equations developed using method 4 gave better results suggesting the adaptation of the methodology to the studied FA. It confirms by the obtained validation coefficients of determination. Highest values were observed for the equations with the highest RPD values except for C18:0. The ability to predict FA using method 4 gave superior results to those shown in previous publications.
