Espectroscopia de reflectância no infravermelho próximo de fezes para predizer variáveis nutricionais de vacas leiteiras confinadas

Abstract

The individual intake of nutrients, as well as variables related to this characteristic are essential within a system of precision nutrition and yet difficult to measure individually when you have group-raised animals. It has been proposed that the chemical information contained in the faeces is representative of the diet consumed and is related to ingestion and digestibility. Near infrared spectroscopy (NIRS) was used to analyze 234 faecal samples from 64 lactating dairy cows fed with TMR, from 5 experiments carried out under similar conditions, in which the individual CMS measured in a tie-stall, digestibility and composition of the consumed diet (CP, NDF and starch) and food selection (Sorting starch and NDF). Two softwares were used for the statistical analysis of the data, the software The Unscrambler and Python. The spectra were analyzed with partial least squares regressor (PLSR), support vector machine regression (SVR), K- nearest neighbor regression (KNNR) and gradient boosting regression (GBR) with or without pretreatment of the set of data (MSC or SNV). Two validation strategies were tested: remove from the training set all data points of an animal (leaving an animal out - LOAO) or an experiment (LOEO) and use the excluded data set to test the algorithm. The results demonstrate that the most used analysis for NIR spectra (PLSR with pretreatments) was not adequate to deal with the complex interaction existing between the analyzed parameters and the faecal spectra. No major differences were found between the algorithms, the best algorithm varied according to the variable studied. The best validation strategy was LOAO. To evaluate the models, we used the RMSE. The CMS showed an error of 2.98 kg / d, accompanied by values below 0.5% for the composition of the consumed diet. For digestibility, RMSE values below 8.71% were obtained. The best accuracy for particle selection was the 8mm sieve, with an average error of 8.85%. For EUN the average was 2.65% and for EA equal to 0.23 Kg milk / Kg DM. As conclusion the analysis of spectral data is influenced by several external factors that will determine which method is best to be used to obtain better results. It is important to note that, although the results are promising and the validation strategies are aimed at minimizing biological entanglement within a cow or assay, the data set came entirely from the same herd. Thus, it is possible that the model is less accurate when used in a completely different herd and environment. It is necessary that these algorithms be explored, looking for the best hyperparameters, using a larger data set and with external validation to verify the applicability of these models.