Selection of methods to analyze body weight and feed intake data used as inputs for nutritional models and precision feeding in pigs

The progress of technologies (sensors, automates) in precision livestock farming enables the development of innovative feeding techniques such as precision feeding of individual animals, potentially applicable in practice. In addition to the design of adapted feeders, precision feeding requires decision support tools to manage data and apply nutritional models that calculate the optimal feed composition and allowance. These calculations require to predict body weight (BW) and feed intake (FI) of individual pigs according to past performance. To select the most accurate prediction method, three statistical methods were tested on a dataset of measurements of BW and FI for 117 pigs: the double exponential smoothing (DES) method, multivariate adaptive regression splines (MARS), and the k-nearest neighbors (kNN) method. These methods were tested in relation to data sampling frequency (i.e., daily or weekly measurements) and data availability. The capacity to predict BW or FI was evaluated through the mean error of prediction. The kNN method appeared suitable if few historical data are available as it requires not more than 3 historical data. The MARS method was better than the DES method to predict daily BW, but the DES method was better in predicting the daily cumulated FI. The DES method seemed also more applicable for weekly BW data, requiring only 3 historical data to make a prediction. These methods could be used for performance prediction in a decision support tool for precision feeding. This study was performed in the Feed-a-Gene Project funded by the European Union’s H2020 Programme (grant agreement no 633531).