Application of unsupervised learning to cluster swine breeding herds based on key performance indicators in Southern Brazil

Abstract Reproductive performance in sow farms has improved over time in most production systems. However, variability among farms remains substantial. This study employed unsupervised analysis to cluster farms based on reproductive indicators and assess the farm-level characteristics associated with performance differences. A total of 22 breeding herds in the Southern Brazil region were surveyed to gather demographic, labor, infrastructure, environmental, management, and production data. Monthly reproductive performance indicators from 2022 and 2023 were collected, including farrowing rate (FR), total piglets born (TPB), piglets born alive (TBA), stillborn (SB), and pre-weaning mortality (PWM). A K-means clustering analysis grouped farms based on five key reproductive performance indicators (KPIs). Cluster identification was used as an independent variable to identify differences in survey responses. The unsupervised model suggested two clusters based on the consensus of 26 indexes. The high-performance cluster (Cluster 2, n = 12) exhibited a higher FR (+3.6%), lower SB (-1.6%), and PWM (-1.7%) compared to the low-performance cluster (Cluster 1, n = 10) (P ≤ 0.01). No significant differences were found between clusters in labor characteristics (P ≥ 0.24). However, the high-performance cluster had a lower percentage of farms using farrowing induction in gilts, a reduced percentage of re-serviced sows included in the breeding groups, fewer farrowing crates per room, and a higher rate of farms introducing creep-feeding earlier during lactation (P ≤ 0.05). This study demonstrates the effectiveness of an unsupervised clustering approach in classifying breeding herds based on five KPIs. This method can identify key farm-level factors that distinguish herd performance, offering insights for improving swine reproductive performance.