Disease detection in pigs based on feeding behaviour traits using machine learning

Abstract

Disease detection is crucial for timely intervention to increase treatment success and reduce negative impacts on pig welfare. The objective of this study was to monitor changes in feeding behaviour patterns to detect pigs that may need medical treatment or extra management. The data included 794,509 observation days related to the feeding behaviour and health information of 10,261 pigs. Feeding behaviour traits were calculated including the number of visits per day (NVD), time spent in feeding per day (TPD), and daily feed intake (DFI). The health status (sick or healthy) of pigs were predicted based on the features including the original feeding behaviour traits and features derived from those using a machine-learning algorithm (Xgboost). The predictions were based either on the features from the same day (one-day window), from the same day and two previous days (three-day window), or from the same day and six previous days (seven-day window). The model based on the seven-day window gave the most robust results and achieved an 80% AUC, 7% F1-score, 67% sensitivity, 73% specificity, and 4% precision. The analyses indicated that the features related to the deviation of a pig's observed TPD and DFI from the expected TPD and DFI were the most informative, as they gained the highest importance score. In conclusion, the feeding behaviour-based features gave good sensitivity and specificity in predicting sickness. However, the precision of the method was very low, possibly due to low prevalence of the monitored sickness symptoms, limiting the application of the approach in real-life.