Development and evaluation of predictive models for real-time decision making on dairy farms
Publication Date
July 23, 2026
Creator
Glover, Ian
ECBHM, Quality Milk Management Services Ltd., University of Nottingham
Abstract
Global demand for dairy products is forecast to rise over the coming decades in response to human population growth. Technological innovations are important for ensuring that this demand is met in a socially, environmentally and economically sustainable manner. Predictive models are a cornerstone of precision livestock farming, leveraging data collected on farms for making accurate predictions regarding the health or welfare of livestock. Model predictions can be incorporated into evidence-based, real-time decisions, informing management changes or disease treatments to optimise sustainability. The aim of this thesis was to develop and evaluate such predictive models for use in real-time decision making by farmers or their advisors.
Data pertaining to two prominent diseases of dairy cattle, paratuberculosis and mastitis, were utilised in this thesis. In Chapter 2, a model was developed which predicts the probability of cure of a case of clinical mastitis. Such predictions are useful for making treatment decisions. For example, cows with a sufficiently low cure probability can be selected for non-antimicrobial treatment. The model in Chapter 2 is noteworthy for the absence of a predictor variable indicating the species of bacterial pathogen; predictors in the model include only cow- and herd- level variables and do not require pathogen-identification.
In Chapters 3 and 4, modelling was used to improve interpretation of serial individual cow antibody ELISA results collected within paratuberculosis control-schemes. Interpretation of a series of multiple results, as opposed to a single test result, is non-trivial, and is likely to lead to misclassification of cows with regards to risk of paratuberculosis infection. In Chapter 3 repeated Bayesian updating was investigated, using a likelihood ratio derived from a i pair of Bayesian models to update the posterior probability of infection at each paratuberculosis test result. The poor calibration of the posterior probabilities in Chapter 3 prompted the application of unsupervised learning (hidden Markov models) to longitudinal ELISA result data in Chapter 4, yielding a model which assigns cows to one of a number of ordinal latent states. The latent states are likely to represent different stages of disease, and can provide an early warning of disease progression. Of particular interest in Chapter 4 was the disclosure of a group of cows (state 3) whose test results would usually be considered “negative” according to conventional interpretation of antibody ELISA results, and yet are at heightened risk of progressing to positive test results. In Chapter 5, the novel latent states were compared with the framework most commonly employed in the United Kingdom (the “J-status”) for interpreting serial paratuberculosis ELISA results. State 3 cows were found to be often concurrently low-risk according to this framework. Furthermore, predictions from the hidden Markov model developed in Chapter 4 were informative in a model created in Chapter 5 that predicts the probability of the onset in the subsequent twelve months of the highest-risk category (“J5”) according to the conventional paratuberculosis risk framework. These predictions are expected to be useful in making pre-emptive management decisions (for example breeding decisions) according to cows’ forecasted risk of progression to “J5” status. In Chapter 6, the implications of the latent states were studied further using inferential models which quantified the association of the latent states with 305-day milk yield, risk of service and risk of conception. Cows in the highest latent states were found to have reduced milk yield compared to those in the lowest states, corroborating the hypothesis that the latent states are aligned with disease progression.
In this thesis, some novel statistical techniques were explored, yielding practically-useful validated prediction models for real-time decision making, and advancing knowledge of the analysis of routinely-collected low-dimensional farm data.
Item Type
ethesis
Thesis Type
PhD
Supervisors
Bradley, Andrew
University of Nottingham
Green, Martin
University of Nottingham
O'Grady, Luke
University of Nottingham
Subjects (LC)
Associated Schools / Departments
Related Publication URL(s)
UoN Repository URI
Except where otherwise noted, this item's license is described as
File(s)![Thumbnail Image]()
Name
Ian Glover PhD Thesis.pdf
Size
3.84 MB
Format
Adobe PDF
Checksum (MD5)
67e330e81b86787d0e0059bbda26c97e