Hannon, Fergus
(2025)
Inferential and predictive modelling of transition success in dairy cows on automatic milking systems.
PhD thesis, University of Nottingham.
Abstract
The health and welfare of the transition cow impacts the economic, environmental and social sustainability of the dairy industry. Despite significant improvements in our understanding of the physiological challenges this period poses, limited improvement in the morbidity and mortality rates associated with transition have been reported in the
previous two decades. Transition cow monitoring programs (TMPs) are a commonly advocated means of reducing the impact of poor transition health. To date, such programs have been largely diagnostic in nature, focusing on the detection of specific disease states utilising labour intensive monitoring techniques. Automatic milking systems (AMS) offer an opportunity to develop fully automated monitoring systems which may be applied using a prognostic rather than diagnostic approach.
Prognostic TMPs provide dairy producers with predictions relating to long-term performance outcomes which may be used to facilitate pre-emptive management practices aimed at preventing or mitigating the losses associated with poor transition health. The aim of this thesis was to investigate the relationship between production and behaviour data
as collected by AMS in the early post-partum period, and subsequent performance assessed using milk production relative to expected, reproductive performance and the risk of removal from the herd in early lactation. An emphasis was placed on the predictive power of this data and its potential utility within a prognostic transition cow monitoring program.
A convenience sample of 46 herds was recruited on a voluntary basis from the UK and Republic of Ireland. Criteria for inclusion was the use of a Lely Astronaut milking robot under free-flow traffic conditions in conjunction with rumination and physical activity monitoring technology. Production variables analysed relating to milk quantity and quality included milk yield, milk yield acceleration, fat and protein content as well as conductivity. Behaviour parameters available via AMS included the number and nature of cow-robot interactions. In addition to this, auxiliary data sources including daily rumination and activity parameters, recorded using a neck mounted accelerometer and, historical cow-level production data as recorded by farm management software were also available. Within this thesis four outcomes were investigated. Chapters 3 and 4 explore the relationship between AMSdata collected over days 1-3 post-partum and Yield Deviation, defined as the disparity between recorded milk production and expected milk production over the first 30 days post-partum. Chapters 5 and 6 examine the relationship between data collected from 1-21 days in milk (DIM) and two reproductive outcomes, Expression of Oestrus or Insemination (EOI); The recording of an oestrus or insemination event between DIM 22 and 65, and Conception to First Insemination (CFI), the conception rate to a first insemination between DIM 22 and 80. Finally, Chapter 7 examines data collected over days 1-3 post-partum and subsequent survival using removal from the herd by 100 DIM.
Mixed-effect multivariable models reported in Chapters 3, 5, and 7 serve to quantify the statistically significant associations between AMS production and behaviour data, and their respective outcomes while accounting for the random effect of herd and confounding variables.
The development, and external validation of machine learning models for the prediction of production and fertility outcomes, described in Chapters 4 and 6 respectively, assess the degree to which AMS data,in combination with auxiliary data sources, may be leveraged into
meaningful improvements in animal health through predictive TMPs. Chapter 6 also examine the marginal effects of auxiliary data sources on model performance by assessing the accuracy with which AMS data can predict reproductive performance with and without rumination, activity and historical production data. Chapter 7 provides a direct comparison of mixed-effect inferential modelling and machine learning predictive models for the odds of removal from the herd by DIM 100 using AMS production and behaviour data in isolation.
In Chapters 3 and 7 we demonstrate that AMS production and behaviour data collected overs days 1-3 post-partum has significant association with Yield Deviation at 30 DIM, and the risk of removal from the herd by 100 DIM. Likewise, in Chapter 5, data collected prior to day 22 post-partum demonstrated significant association with reproductive outcomes up to 80 DIM. Across all outcomes, variables relating to milk yield, rate of milk yield acceleration and fat-to-protein ratio were found to be statistically significant. These associations highlight the transition period, and in particular days 1-3 post-partum as a critical inflection point within the lactation cycle. Furthermore, it demonstrates the potential for AMS sensor data collected during this time to be
incorporated into a prognostic TMP. However, the coefficient of determination attributed to the fixed effects within the final models for both reproductive and survival outcomes were found to be low, indicating that the explanatory power of these variables is limited.
Assessed in Chapters 4, 6, and 7, transition cow data demonstrated moderate group level-predictive power for Yield Deviation at 30 DIM, and reproductive outcomes EOI and CFI, but failed to demonstratepredictive power for the risk of removal by 100 DIM. The predictive power of AMS and auxiliary data sources examined in Chapters 4 and 6 represents a critically important finding in support of the premise of prognostic TMPs. While predictive performance is moderate, these findings highlight the potential utility of this data to identify animals likely
to experience poor production or fertility performance in the early stages of lactation and should encourage further investigation of how this data may be applied within TMPs. However, the absence of predictive power for the risk of removal in early lactation, reported in Chapter 7 highlights a potential limitation of this approach to transition cow monitoring, particularly where the lag between observations and outcomes is prolonged. These results also serve to demonstrate the risks in the use of inferential models to imply predictive power and the need for externally validated predictive models to be incorporated into the assessment of the potential utility of novel data sources. The failure to demonstrate a statistically significant increase in model performance following the incorporation of auxiliary data sources, as reported in Chapter 6, highlights the challenges of balancing model accuracy with generalisability and ease of deployment in an environment of rapidly increasing data complexity.
The work presented within this thesis examines a novel means of transition cow monitoring, one which seeks to assess transition health using subsequent production, fertility and survival outcomes. The inferential models reported demonstrate significant statistical association between AMS data and each outcome of interest. However, the predictive power of this data remains limited when applied at the level of the individual, particularly as it relates to the risk of removal within the first 100 days post-partum. Despite this, group level classification of milk production and fertility outcomes demonstrated potential for incorporation into prognostic TMPs. This represents a critical advancement in the field of transition cow monitoring and may
offer an effective means to improve the health of transition cows and hence, the sustainability of the dairy industry.
| Item Type: |
Thesis (University of Nottingham only)
(PhD)
|
| Supervisors: |
Laura, Randall
Martin, Green
Chris, Hudson
Luke, O'Grady
Anneke, Gouw |
| Keywords: |
Cows; Behaviour data; Milk production; Reproductive performance; Transition cow monitoring; Predictive data |
| Subjects: |
S Agriculture > SF Animal culture |
| Faculties/Schools: |
UK Campuses > Faculty of Medicine and Health Sciences > School of Veterinary Medicine and Science |
| Item ID: |
81053 |
| Depositing User: |
Hannon, Fergus
|
| Date Deposited: |
24 Jul 2025 04:40 |
| Last Modified: |
24 Jul 2025 04:40 |
| URI: |
https://eprints.nottingham.ac.uk/id/eprint/81053 |
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
