Comparison of feed tables, empirical models and near-infrared spectroscopy to predict chemical composition and net energy of pelleted pig feeds

The determination of the net energy (NE) of a pig feed requires expensive in vivo experiments. Alternatives are available whether the feed ingredient composition is known or unknown. We compared three estimation methods: feed tables, empirical models based on chemical parameters and in vitro digestibility and near-infrared spectroscopy (NIRS) calibrations based on feed and feces spectra. All methods were evaluated for their accuracy in predicting chemical composition with a focus on NE, the applicability in practice, and practical limitations. From a total dataset of 80 feeds and 480 feces samples originating from 3 in vivo digestibility trials, 28 feeds were used to examine the use of feed tables, 62 feeds for empirical models and 62 feeds and 310 feces samples for NIRS calibrations. Feed tables accurately predicted most chemical components (R2 > 0.90), except for crude ash, sugar, acid detergent lignin and moisture. The dry matter (DM) content was underestimated on average by 16 g/kg and 17 g/kg from Dutch (CVB) and French (INRA) tables, respectively. The direct use of tabular NE values overestimated feeds containing beet pulp. The prediction of NE, calculated according to the CVB system, from the tabular chemical composition and digestibility coefficients of the feed ingredients resulted in a standard error of estimate (SEE) of 0.29 and 0.43 MJ/kg for CVB and INRA tables, respectively. Expressing these values on DM basis decreased SEE to 0.22 and 0.30 MJ/kg DM for CVB and INRA, respectively. The best empirical model to predict NE based on chemical composition resulted in a standard error of cross-validation (SECV) of 0.21 MJ/kg. Incorporating the in vitro digestibility of organic matter allowed to decrease SECV to 0.18 MJ/kg. A NIRS calibration based on feed spectra alone resulted in a SECV of 0.33 MJ/kg, whereas the combination of feed and feces spectra decreased SECV to 0.26 MJ/kg. All three estimation methods have advantages and disadvantages. Feed tables allow a quick estimation of the composition and average NE content, but they can only be used for feeds with known and common ingredient composition. Empirical models allow the most accurate estimate of the NE of feeds of unknown composition, but they require expensive and time-consuming analyses. Finally, a NIRS calibration based on feed and feces spectra enables a fast estimation of NE, but is less precise than empirical models and requires sample preparation and a large dataset.