Accurate and interpretable prediction of poor health in small ruminants with accelerometers and machine learning

Abstract

Assessment of the health status of individual animals is a key step in the timely and targeted treatment of infections, which is critical in the fight against anthelmintic and antimicrobial resistance. The FAMACHA scoring system has been used successfully to detect anaemia caused by infection with the parasitic nematode Haemonchus contortus in small ruminants and is an effective way to identify individuals in need of treatment. However, assessing FAMACHA is labour-intensive and costly as individuals must be manually examined at frequent intervals. Here, we used accelerometers to measure the individual activity of extensively grazing small ruminants (sheep and goats) exposed to natural Haemonchus contortus worm infection in southern Africa over long time scales (13+ months). When combined with machine learning, this activity data can predict poorer health (increases in FAMACHA score), as well as those individuals that respond to treatment, all with precision up to 83%. We demonstrate that these classifiers remain robust over time. Interpretation of trained classifiers reveals that poorer health significantly affects the night-time activity levels in the sheep. Our study thus reveals behavioural patterns across two small ruminant species, which lowcost biologgers can exploit to detect subtle changes in animal health and enable timely and targeted intervention. This has real potential to improve economic outcomes and animal welfare as well as limit the use of anthelmintic drugs and diminish pressures on anthelmintic resistance in both commercial and resource-poor communal farming.