Learning patterns from time-series data to discriminate predictions of tail-biting, fouling and diarrhoea in pigs

The occurrence of tail-biting, fouling and diarrhoea affect the welfare condition of commercially farmed pigs, and also reduces the profitability and sustainability of commercial pig farms. Early detection of these unwanted events can be achieved by detecting behavioural changes at an early stage in the animals. While pigs are likely to change their diurnal behaviour when their health and welfare conditions are compromised, the early detection of such problems can be a challenging task due to the stochasticity in pigs' behaviour. In this study we proposed using a stacked bidirectional long short-term memory (LSTM) and feedforward neural network architecture in order to detect behavioural changes through three separate different dynamic models. Data on pen level was collected on pig water consumption, pen level temperatures, indoor climatic data (ventilation, cooling, heating, and relative humidity) from a total of 112 pens, giving a total of 7632 samples were used to train and test the proposed neural network. To reduce training bias, the network was trained using a stratified 10-fold cross validation approach. The performance on the test set was measured using the Area Under the curve of the Receiver Operating Characteristic (AUROC). Using a 7-day window, an AUROC of 0.782, 0.775 and 0.820 was obtained for prediction of tail-biting, diarrhoea, and fouling. This study demonstrated that the proposed neural network architecture can successfully learn the behavioural changes that causes specific welfare and health problems in pigs. Besides the ability to effectively learn to robustly predict a range of health and welfare problems, the approach taken in this study did not involve the laborious task of manual feature engineering that traditional machine learning often require. Automatically learning complex relationships from temporal data can in the future speed the process of model development for health and welfare problem detection. However, the features that neural network models learn are often abstract, and difficult to interpret. For more interpretability, future studies should investigate how machine learning techniques perform using a range of manually extracted features from temporal data.