AI to monitor changes in social behaviour for the early detection of disease in dairy cattle

In the UK, dairy milk is a key part of the economy and an important source of nutrition. There are several diseases that regularly develop in UK dairy cows which compromise health and welfare, and lead to economic losses for the farmer and industry. Ill cows have also been found to contribute disproportionately to methane emissions and hence the environmental sustainability of the sector. In addition, high welfare is more important than ever to satisfy societal demands for food production.

To help farmers detect and treat these diseases, numerous solutions for automated monitoring of dairy cattle are now available to farmers. A critical disadvantage of all these technologies is that they are focussed on detecting the observable symptoms of later stage disease, when treatment options may be limited, reduction of milk production persistent and animal welfare more severely compromised.

A cow's response to infection and trauma is to de-prioritise behaviours not immediately essential to survival and recovery - such as social interactions - in favour of those that remain critical for longer, In a recent study we have found that social exploration, the grooming of others and receiving headbutts were lower in individuals with early stage mastitis. We hence hypothesise that social behaviour changes could be early predictors of disease.

Detecting social behaviour changes is difficult for the busy farmer, but is possible by monitoring them at key focal points, such as when queueing for milking or feeding at the feed bunk, using video cameras and artificial intelligence (AI). We have developed highly robust AI that can track the motion of cows in video and recognises each individual through their distinctive coat pattern. Others have now demonstrated good classification of affiliative and agonistic social interactions from video and hence we now propose combining the two ideas to track changes in activities and social behaviours over time for each identified cow in a herd. From collecting two years of video from 64 cameras covering the main barn at our John Oldacre Centre dairy farm, we will train a model that learns what types of behaviours change over time that are indicative of different early stage diseases. We will focus on mastitis and lameness, as these diseases have the greatest incidence in our data and are the most important for the UK dairy industry. At the same time, we will sample the saliva of a subset of our herd so we can determine general levels of inflammation, enabling us to see how specific our behavioural predictors are to particular diseases.

Dairy farmers are specialists in the behaviour and personalities of their cattle and their input will be vital to helping understand vagaries in farm data and how our system is functioning. We will test our system by deploying it at a network of recruited farms, and will conduct in-depth semi-structured interviews with the farmers regarding their experiences of camera placement (including intrusiveness and social acceptance by farm workers), operation and any other perceived impacts to their farms, farm workers or animal management, health and welfare.

It is also critical that we design the system with all facets of industry, to engage their diverse insights and expertise in setting alert levels, designing user-friendly interfaces that will be well placed to be uptaken and discussing additional routes to market such as for disease surveillance. We have therefore assembled a consortium of partners covering all key areas from farmers to vets, the supply chain, data/diagnostic service providers and business development, all of whom we have a proven track record of successful engagement and impact with. Through consultation we will develop a sustainable strategy for meaningful lay stakeholder and public involvement with our system and results, helping to promote a widespread understanding and public/stakeholder acceptance of the system.

Spotting the early, subclinical stages of endemic disease in individual cows is essential for successful treatment, maintaining high welfare standards and sustainable intensification. A critical disadvantage of current technologies for monitoring cattle is that they are focussed on detecting the physical manifestations of later stage clinical disease. Based on behavioural research and our recent work, we hypothesise that social behaviour changes are early predictors of subclinical disease. In this project, we will harness our series of published advancements on computer vision-based individual cow tracking and re-identification together with deep learning activity classifiers of different behaviours and types of social interaction to develop early prediction models for mastitis and lameness based on changes in social behaviour. The Bayesian prediction models will be trained using two years of comprehensive video capture of the whole herd at our John Oldacre Centre, together with fortnightly somatic cell count measurements, mobility scoring and saliva testing for determining generalised inflammatory response.

This project includes a major focus on farmer and industrial co-design, including eight partners covering vets, the supply chain, diagnostic services and business development. A network of external testbed farms will be recruited to optimise and fully validate the translational potential of the system for prediction generalisability and with farmers regarding acceptability and usability. Qualitative data from interviews, observed one-on-one interactions and farmer and stakeholder focus groups will be thematically analysed to identify resource feasibility and potential impact within the industry. Pathways to impact include development towards a commercial solution for herd health management, deployment for farm assurance and disease surveillance, and as a platform underpinning next-generation animal health and welfare research.