FARM interventions to Control Antimicrobial ResistancE

The main FARM-CARE project focuses on exploring techniques for controlling the spread of antimicrobial resistance (AMR) in pig farms. Pigs are the main contributor to antimicrobial use (AMU) in farm animals and a recognized source of AMR to farm workers, the community and the environment. The project builds upon the established notion that stress is a major driver for pig disease, which in turn contributes to AMU. Our hypothesis is that AMR spread can be controlled by firstly limiting the common practice of mixing, which is a recognized cause of stress and disease in pigs, and secondly, by applying biosecurity measures to prevent AMR transmission to farm workers and the community. We aim to understand the impact of these two interventions on AMU and AMR reduction, and to develop two complementary interventions based on machine learning. The project will assess the cost-effectiveness of each intervention from the perspective of the farmers, the environment and the wider society. The UK team will focus on detecting high-risk piglets using facial imaging and exploring patterns in metagenomics and metadata associated with post-weaning diarrhoea.

FARM-CARE is a multi-actor project involving three higher education institutions, three public organisations, and one private company with established national and international networks that will facilitate dissemination and stakeholder involvement. Each partner brings one or more areas of expertise with clearly defined roles and responsibilities. Participation by an LMIC partner, that represents over 90% of pig farmers in Colombia, will allow assessment of the farm interventions in a different production setting and offer capacity building in this country.

The project will develop innovative machine learning tools to identify predictors of stress and disease in new-born piglets. We will assess the relative merits of the different interventions from the perspective of the farmer and society using business case and cost-effectiveness analyses.

Two different machine learning approaches will be used to identify high-risk piglets predisposed to Post-weaning diarrhoea (PWD) and other diseases with the intention of being able to separate them from healthy individuals and/or reduce their occurrence by breeding management. The first approach is based on facial feature analysis of stress in sows, whereas the second is based on integrated analysis of faecal metagenomics data and metadata collected longitudinally from individual piglets. In both approaches, data will be analysed retrospectively to identify behavioural and metagenomics markers that can be used to predict PWD and other diseases of interest.

We will use machine learning to detect facial expressions in sows from image data captured during gestation and around farrowing to estimate the potential stress levels (as a percentage likelihood that a pig is stressed versus not stressed) the sows are experiencing. These estimates will then be correlated with their stress biomarkers and disease measurements (e.g. faecal scores) taken from their piglets. Similarly, metagenomics data will be integrated with collected metadata and analysed using a machine learning approach. A statistical model will be used to check for overdispersion and collinearity. The outcome of this task will be a score for how strongly these different measures are able to predict disease.

The collaborative work on automated expression detection for biometrically identified animals, also has important wider application for observation and wellbeing monitoring of wildlife in the study of biological processes - a key component of NERC's remit in the UK.