TY - THES T1 - The prediction of energy balance of dairy cows from animal, feed, and milk traits with special regard to milk fatty acids A1 - Becher,Vera Y1 - 2017/04/24 N2 - The objective of the present study was to predict the energy balance (EB) of dairy cows from animal, feed, and milk traits. As the milk fatty acid (FA) profile is known to react to physiological conditions like an energy deficit, special regard was given to milk FA in order to identify new potential indicators for negative EB. Visiting six experimental stations in Germany, single milk samples were taken from dairy cows between their 6th and 133th day in milk to create a dataset covering a large spectrum of EB and a variety of practical diets. The milk composition was analyzed by mid-infrared spectrometry, and the milk FA profile via gas chromatography. Energy balance (MJ NEL/d), as response variable, was calculated by subtracting the cow’s energy requirements from energy intake. As candidate variables parity, day of lactation, dietary nutrient composition, milk yield, milk composition, and the milk FA profile were provided resulting in a pool of 62 potential predictors. The prediction of EB was performed in two different ways: first, an automated stepwise variable selection was performed with the whole variable pool (GLMs-N) and with FA only (GLMs-FA-N). As this method recently earned criticism, some other methods were also tested for a first variable selection: the regularized linear regression models Lasso, elastic net (ENET), adaptive Lasso (AdaLasso), and adaptive elastic net (ADAENET). As a machine learning method which also considers interactions and non-linear relationships random forests were also applied. The first variable selection was performed using a five-fold cross-validation which resulted in five models per selection method. All chosen effects were combined to one model (MODEL1) for each method, respectively. Following this, the individual effects of the MODEL1 were used for a forward selection based on the corrected Akaike Information Criterion (AICC) for further model reduction, resulting in MODEL2. Then, the non-significant effects were removed from the MODEL2, achieving the final MODEL3 for each method. The final models were validated using leave-one-out cross-validation. The models showed adequate correlations (r) between the predicted and the observed EB in leave-one-out cross-validation: although GLMs-FA-N had the lowest accuracy (r = 0.79), the result was still remarkable and showed how much information milk FA alone can provide. GLMs-N and AdaLasso performed best with r = 0.86 and 0.85 containing 21 and 18 predictors, respectively. However, other models like ADAENET achieved only slightly lower accuracy (r = 0.83) with only 6 predictors. The composition of the predictors was relatively similar in all models. All (except for GLMs-FA-N) contained days in milk, milk yield, C18:1c9, C15:0iso, and the ratio of omega-6 to omega-3 FA (n-6/n-3) as effects with the strongest impacts on the prediction. While milk yield, days in milk, and C18:1c9 mirrored physiologically obvious effects, the strong and positive impact of n-6/n-3 and C15:0iso was unexpected. The n-6/n-3 ratio might be physiologically connected to EB as might reflect the dietary forage-to-concentrate ratio which influences dietary energy content and thus EB. The importance of C15:0iso, a FA arising from microbial FA synthesis in the rumen, could not be explained satisfyingly. The nature of the potential physiological connections between EB and some FA like C15:0iso or n-6 or n-3 FA might require further research. The present study showed that it is possible to predict the cow’s EB from animal and milk traits with an adequate accuracy. As long as the diets have similar composition and not contain ingredients which strongly affect the milk FA profile, dietary effects have not to be taken into account. However, a practical application of the obtained models is not yet possible: First, as the dataset was relatively small (n = 248), it is not clear whether or not the models would perform adequately with independent datasets. Second, FA analysis by gas chromatography is very expensive. Third, even if gas chromatographic analysis were affordable for standard milk analysis, there are some highly variable, very low concentrated FA as predictors in the models, which might be prone to laboratory effects, and this could spoil the predictions. Although under criticism, automatic stepwise selection provided the best performing model and thus seems sufficient for practical issues like the one dealt with in the present study. However, the differences in accuracy between the applied methods were very small and as regularized linear regression methods, especially ENET and ADAENET, are supposed to deal better with highly correlated variables, it might be safer to use them with datasets containing highly correlated variables such as the one used in the present work. KW - Milchkuh KW - Energiebilanz KW - Milch KW - Fettsäuren CY - Hohenheim PB - Kommunikations-, Informations- und Medienzentrum der Universität Hohenheim AD - Garbenstr. 15, 70593 Stuttgart UR - http://opus.uni-hohenheim.de/volltexte/2017/1347 ER -