Prediction of litter performance in lactating sows using machine learning, for precision livestock farming

Abstract

• The best predictions were obtained with farm-level training and ensemble methods. • Reliable predictions were also obtained with simple machine learning algorithms. • Effects of farm, litter size, and parity on litter performance were confirmed. • The presented approach enables the estimation of milk nutrient output in sows. Predicting litter performance in lactating sows is an essential step towards the development of decision support systems for precision feeding in lactating sows. Numerous factors affecting litter performance have been described in literature. However, predictive models working on-farm in real time are not available. The main objectives of this research was to (i) explore 4 different machine learning strategies, and (ii) identify the best supervised learning algorithm in order to obtain reliable predictions of litter performance. This study was carried out with data obtained from 6 experimental farms over the last 20 years. Algorithms were trained to predict the litter weight at weaning using a set of 4 numeric and 3 categorical features, and a method for predicting secondary litter performance and nutrient output in milk from the predicted litter weight at weaning was evaluated. To evaluate the reliability of predictions within each farm, the mean error per farm (ME f ) and the mean absolute percentage error per farm (MAPE f ) were computed. The best performance for the prediction of litter weight at weaning was obtained with an ensemble algorithm with farm-level training and testing (ME f = −0.14 kg; MAPE f = 9.01%), but performance with simple linear regression was very close (MAPE f = 9.30%). Learning across all farms only achieved comparable results with the neural networks algorithm, but at higher computational costs. The method for predicting secondary litter performance and nutrient output from the predictions of litter weight at weaning reveals that the ME f remains close to 0, and that the MAPE f only increases by a few percentage points. This study confirms the effect of numerous factors known in the literature to affect litter performance, such as litter size and parity of sows, but also revealed huge variations between farms. According to this study, reliable predictions could be obtained with interpretable supervised algorithms trained at farm level, with features that can be easily measured on-farm. This study thus shows that on-farm data are necessary to accurately train models and make reliable predictions at farm level. These predictions could be used by decision support systems in order to develop precision feeding approaches in lactating sows.