Employing AI-assisted automated monitoring system and social network analysis for quantifying social structure and detecting harmful behaviors in pigs

Negative social behaviors represent a welfare and economic problem for farm animals worldwide. The time-consuming nature of observing behavior on a large scale means that social problems are not visible until their negative physical effects are advanced. However, identifying the key initiators or propagators of fighting and/or biting may not be possible by observing injury alone. Therefore, in this study, we investigated the possibility of constructing and analyzing social networks from AI-assisted automated monitoring data in pigs and the feasibility of using spatial proximity association as an indicator of harmful social interactions. Data were collected using automated recording systems that captured 2D camera images and videos of 6 pens of pigs (16-19 per pen) on a PIC breeding farm (USA). The system records continuous video footage with the associated real-time ear-tag ID, elapsed time, posture (standing, lying, sitting) and XY coordinates of the shoulder and rump for each pig. The validation of automated identity, posture and location records show 97-100% agreement with human observations (Agha et al., 2024). Pig movements were recorded from 10:00 to 19:00h for 6 days; 3 days immediately after regrouping and 3 days, 60 days after regrouping. Proximity data was used to create weighted social networks. Group level metrics (degree, betweenness, and closeness centralization) significantly increased from the early to late growing periods (p< 0.02), highlighting that inequality in proximity between pigs increased over time. Largest clique size remained unchanged (p=0.28), but the number of maximal cliques (fully connected subgroups) significantly decreased from the early to late growing period (p=0.007). Individual SNA traits were mostly stable over these periods. Measuring the behavior of prominent individuals during the time they are in proximity will allow targeting of management interventions to improve welfare outcomes. We tested an initial dataset as proof of concept that proximity signatures can be used to identify aggression. Video footage of two pens of 19 pigs were observed and 37 mutual fighting bouts were identified. Proximity matrices were calculated for each pen using shoulder XY location, for the duration of the bout, lasting 2-42 s (median 5.0s). 81% of fighting dyads were identified as in proximity (< 0.5m) compared with 7% (468/6290) of non-fighting dyads, χ2=7.0, p=0.008. Of the dyads in proximity, fighting dyads were in proximity for over twice as much of the interval (median: Q1-Q3); (92: 31.4-96.1%) as non-fighting dyads (44.3: 16.8- 86.1%), p< 0.001 Refinements to improve sensitivity and specificity to diagnose and characterize aggressive behavior from larger samples of proximity data are ongoing. This study demonstrates that integrating SNA with automated data reveals novel insights into pigs’ social interactions and identifies a signature of aggressive encounters using proximity. That could offer promising applications in breeding and management of farmed animals.