Predicting fibre digestibility in Holstein dairy cows fed dry-hay-based rations through machine learning

Calculating the requirements and predicting the feed digestibility are essential to building robust dairy cattle rationing programmes. In the field, a huge number of in vivo observations are needed to develop accurate equations and reliable predictions. The aim of this study was to develop an equation to estimate total-tract potentially digestible NDF digestibility (TTpdNDFD) for lactating cows fed hay-based rations. Individual data from 11 studies, 69 cows, 35 different treatments, and 1 614 observations were included in this study. To develop the prediction equation, the following traits, descriptors of the total mixed ration, were used: ash, starch, CP, NDF, acid detergent fibre, acid detergent lignin, undegradable NDF and potential degradable NDF. Before building the equation with bidirectional stepwise selection in the JMP software, outliers were removed and multicollinearity was checked for all the predictors of fibre digestibility. The model was trained with 10-folds cross-validation. Results showed an R2 of 0.91 and 0.90, and a RMSE of 2.99 and 3.26 in the model for training and validation, respectively. The promising performance of the model suggested that, the fibre digestibility in lactating dairy cows fed dry-haybased rations can be accurately predicted in advance just by using the diet characteristics. From the obtained equation, we predicted the weight and slope of the included covariates, and outcomes confirm that in general the TTpdNDFD is reduced as dietary starch and fast-fermentable fibre increase. This study found that the equation extracted from a neural network, when combined with precision farming techniques, can improve the management of lactating cows and optimise feed planning, monitoring, and cost. It can be used in areas where silages are not used in rations. This provides evidence that accurate equations can be developed from historical data for precision feeding implementation. Further research is needed to expand the dataset and develop equations that can be applied on a large scale. Improving accuracy would involve incorporating representative data from other areas with similar diets into the training set.