Development and evaluation of models for predicting reproductive performance in large commercial dairy herds.

摘要

Machine learning methods, including decision trees, Bayesian networks, and instance-based algorithms, were tested for the development of predictors of conception at first service and status at 150 days postpartum (pregnant or not) in dairy cattle, using more than 300 explanatory variables. The alternating decision tree algorithm, which provided the most useful features and performed well in 10-fold cross-validation, was chosen for the final analyses. This algorithm is capable of handling highly correlated variables with many dependencies and interactions, is robust to missing values and coding issues, and builds very readable trees. By evaluating these variables jointly in a single analysis, and by including variables that had not been examined previously, a better understanding of the relative importance of a broad range of herd management variables and their interactions, with respect to reproductive performance, was developed.; Environmental factors that lead to an increased risk of peripartum diseases, including large, overloaded pre- and postpartum pens, sub-optimal strategies for moving cows among pens, and inadequate facilities had a great impact on fertility. Specific herd management factors, such as frequency of maintenance hoof trimming, timing of initiating BST injections after calving, and length of the voluntary waiting period were also highly associated with fertility traits. Interestingly, several of these explanatory variables (some never evaluated previously) showed a greater impact on fertility than temperature and milk production and may re-direct the allocation of resources in future research and field applications. The impact of body condition score on first service conception (evaluated in a separate study) was much greater than for second and later services; it showed a steep increase in probability of conception as scores went from 2.25 to 3.00 and plateaued thereafter. The predictors described above were developed for field applications for generating probabilities of conception or becoming pregnant by 150 days, but variables included in the resulting models can also be individually evaluated for their impact on these fertility traits.