Short communication: Use of genomic and metabolic information as well as milk performance records for prediction of subclinical ketosis risk via artificial neural networks

• Published in: Journal of dairy science Vol. 98; no. 1; p. 322
• Main Authors: Ehret, A, Hochstuhl, D, Krattenmacher, N, Tetens, J, Klein, M S, Gronwald, W, Thaller, G
• Format: Journal Article
• Language: English
• Published: United States 01-01-2015
• Subjects: 3-Hydroxybutyric Acid - blood; Animals; Asymptomatic Infections; Cattle - physiology; Cattle Diseases - diagnosis; Cattle Diseases - metabolism; Female; Genomics; Ketosis - diagnosis; Ketosis - metabolism; Ketosis - veterinary; Lactation - physiology; Metabolomics; Milk - metabolism; Neural Networks (Computer); Postpartum Period; Risk; prediction; artificial neural network; milk metabolite; ketosis
• ISSN: 1525-3198
• DOI: 10.3168/jds.2014-8602

Subclinical ketosis is one of the most prevalent metabolic disorders in high-producing dairy cows during early lactation. This renders its early detection and prevention important for both economical and animal-welfare reasons. Construction of reliable predictive models is challenging, because traits like ketosis are commonly affected by multiple factors. In this context, machine learning methods offer great advantages because of their universal learning ability and flexibility in integrating various sorts of data. Here, an artificial-neural-network approach was applied to investigate the utility of metabolic, genetic, and milk performance data for the prediction of milk levels of β-hydroxybutyrate within and across consecutive weeks postpartum. Data were collected from 218 dairy cows during their first 5wk in milk. All animals were genotyped with a 50,000 SNP panel, and weekly information on the concentrations of the milk metabolites glycerophosphocholine and phosphocholine as well as milk composition data (milk yield, fat and protein percentage) was available. The concentration of β-hydroxybutyric acid in milk was used as target variable in all prediction models. Average correlations between observed and predicted target values up to 0.643 could be obtained, if milk metabolite and routine milk recording data were combined for prediction at the same day within weeks. Predictive performance of metabolic as well as milk performance-based models was higher than that of models based on genetic information.