235 Unsupervised learning for clustering swine breeding herds based on farm characteristics and management and their impact on reproductive performance in southern Brazil breeding herd

Abstract Introduction: Identifying characteristics that account for variability in reproductive performance among sow farms is challenging through simple field observation. Unsupervised learning effectively clusters farms by their similarities, revealing relationships among various multidimensional characteristics and management practices. This study aimed to classify farms based on farm-level characteristics and identify drivers associated with reproductive performance. Methods: An observational study was conducted across 29 breeding herds from the same company in southern Brazil, examining ten characteristics related to infrastructure and management practices. These included the presence of trees and grass around gestation and farrowing barns, automated control of temperature and lighting in gestation and lactation facilities, barn orientation (east-west), as well as management practices such as encouraging sows to stand during lactation, bump-feeding, and grouping sows by body condition score before moving them into gestation pens. A K-means cluster analysis was performed to group farms based on these variables. The method assessed the homogeneity of characteristics using the Ward-D2 method through a binary distance matrix. The performance data was used to assess differences in performance between clusters. Reproductive indicators were analyzed using the Kruskal-Wallis or ANOVA test. The frequency of farm characteristics between clusters was analyzed using Fisher’s exact test, and all analyses were performed in R software. Results: The model suggested two clusters, Cluster 1 (C1) with 17 farms and Cluster 2 (C2) with 12. The farms averaged 1,269 sows (450–3,024) in C1 and 1,264 sows (330–3,000) in C2. Farrowing rate (FR) did not differ significantly (P=0.21) between Clusters 1 and 2 (90.0% vs. 88.6%, respectively); however, the coefficient of variation of FR was lower in C1 (3.39% vs. 4.31%). No difference in the percentage of return to estrus (P=0.56) was observed between clusters (C1: 5.10%; C2: 5.57%). The total number of piglets born and born alive also did not differ between clusters. However, pre-weaning mortality of piglets was lower in C1 (5.82%) compared to C2 (7.00%; P=0.03). The C1 had a higher percentage of farms with trees and grass surrounding the gestation and farrowing barns than C2 (P< 0.01). Additionally, C1 had a lower percentage of farms that used bump-feeding and encouraged sows to stand during lactation than C2 (P≤0.05). Trees and grass around facilities may affect barn temperature, ventilation, and less exposure to dust and airborne pathogens. Bump-feeding reduces feed intake during lactation. Both factors influence CV of FR and pre-weaning mortality. Sows standing more frequently could increase piglet crushing. Table 1 shows the frequency of farm characteristics across clusters. Conclusion: Cluster 1 demonstrated, on average, better reproductive performance, distinguishing itself from Cluster 2 by having more farms with trees and grass around gestation and lactation barns and fewer farms using bump-feeding or encouraging sows to stand during lactation.