Classification of light Yorkshire pigs at different production stages using ordinary least squares and machine learning methods

• Published in: Animal (Cambridge, England) England), 2024-01, Vol.18 (1), p.101047-101047, Article 101047
• Main Authors: Casellas, J, Salgado-López, P, Lorente, J, Diaz, I Solar, Rathje, T, Gasa, J, Solà-Oriol, D
• Format: Article
• Language: English
• Subjects: Area under the curve; Artificial intelligence; Growth retardation; Live weight; Swine
• ISSN: 1751-7311; 1751-732X
• DOI: 10.1016/j.animal.2023.101047

Pig homogeneity and growth are major concerns for the pig industry today. Variability in pigs' size has a strong impact on profitability as uniformity plays a key role in the overall economic value of pigs produced. This research focused on statistical methods to identify pigs at risk of growth retardation at different stages of production. Data from 125 083 Yorkshire pigs at weaning (18-28 d), 59 533 pigs at the end of the nursery period (70-82 d) and 48 862 pigs at slaughter (155-170 d) were analyzed under three different cut-points (lowest 10, 20 and 30%) to characterize light animals. Records were randomly split into 2:1 training:testing sets, and each training data set was analyzed through an ordinary least squares approach and four machine learning algorithms (decision tree, random forest, and two alternative boosting approaches). A wide range of weighting functions were applied to give increased relevance to lighter pigs. Each resulting classification norm was used to classify light pigs in the testing data set. Both sensitivity and specificity were retained to construct the receiver operating characteristic curve, and the statistical performance of each analytical approach was evaluated by the area under the curve (AUC). In all production stages and cut-points, the random forest machine learning algorithm provided the highest AUC, closely followed by boosting procedures. For weaning BW (WW), factors related to birth BW and litter size accounted for more than 75% of the important prediction factors for light pigs. BW at the end of the nursery period and slaughter BW analyses revealed a similar pattern where WW and BW at the end of the nursery period accounted for more than 40 and 50% of statistical importance among the prediction factors, respectively. Machine learning algorithms are useful tools to easily evaluate the risk factors affecting the efficiency and homogeneity in swine. Since the BW at birth and weaning are key factors, sow nutrition and feeding management during gestation and lactation, along with piglet management during lactation, are identified as important influences on pig weight variability.