Early disease detection for weaned piglet based on live weight, feeding and drinking behaviour

Michel Marcon et al., 70th Annual Meeting of the European Federation Animal Science (EAAP), 26-30 août 2019, Ghent, Belgique, visuels d'intervention

Early disease detection is one of the key to effective disease control in farms and reducing antibiotics usage. A batch of 153 weaned piglets was used to test a first machine learning algorithm in order to predict the individual health state of each animal. In order to build the early disease detection algorithm, nine boxes of 17 piglets has been set up with automata. In real time within this section we knew the number of times each animal went to the drinker or the feeder, the quantity of water and feed it took and its weight. As the golden standard to know either a piglet seems healthy or not, the clinical signs will be observed by trained operators on each pig every workday and recorded on a standardized grid (diarrhoea, cough, lameness…). Then, data collected from this batch of 153 piglets were used to create an algorithm with the software R, based on bagging and random forest machine-learning method. The database was split into learning (70%) and testing (30%). We obtained a global success of 86% of good prediction. 
In order to validate the accuracy of the model, a second batch of 153 piglets was used. Every day, a list of predicted sick pigs was printed automatically, indicating the individual identification of the animal, and its pen. Then, the results of these predictions were compared with the golden standard (observations of clinical signs by trained operators). Out of 3,437 observations (including predictions that the piglet is not sick), the algorithm correctly predicted the status of the piglets 2,462 times. Artificial intelligence has made 72% of good predictions. Regarding the true positive results, 96 alerts out of 117 were actually associated with observations of animals suffering mainly
from diarrhoea within two days (82% of success). Now, the aim is to improve this algorithm in different ways: to test accelerometers to check the activity of each piglet; to be more accurate on recording cough by a microphone (SOMO, Soundtalks); to test if some trajectories of behavioural change are linked to specific diseases (lameness, digestive or respiratory disease) and not only to generic disease. These studies will be part of the Healthylivestock project (EC funded H2020 research project).