Exploring individual variation in gestating sows’ feeding patterns: applying self-organising maps to reveal behavioural types

Large behavioural variation within and among gestating sows hampers distinguishing normal from deviating feeding behaviour. Identifying different behavioural types in sows can contribute to understanding variation and improving the use of electronic sow feeders (ESFs) data for monitoring sow performance and welfare. The aim of this study is to explore the existence of behavioural types in sows’ feeding patterns during the gestation period, focussing on frequency and timing-related feeding traits. For this, self-organising maps (SOMs) were used. A SOM is an unsupervised machine learning method providing ample visualisation support. Data were obtained from in total 1519 sow gestations on two farms (1080 conventional and 439 organic) and were processed and cleaned. Six feeding traits on feed cycle (24 h) level were analysed: 1) sum of nutritive (N-)visits (with feed allowance), 2) sum of non-nutritive (NN-)visits (without feed allowance), 3) hour of the first N-visit (N-start), 4) hour of the first NN-visit (NN-start), 5) interquartile range in hours between NN-visits in one feed cycle (NN-IQR), 6) interval in hours between the last N-visit and first NN-visit (N-NN-range). These traits were analysed separately per farm, thus in total 12 SOMs were trained. Sow gestation patterns were grouped into three clusters per SOM to represent different behavioural patterns during gestation. For three of the traits (N-visits, NN-visits and NN-IQR), one pattern was dominant, with only occasional deviations in individual sows. In contrast, the other three (timing-related) traits N-start, NN-start and N-NN-range showed no clear dominant pattern, suggesting the presence of coexisting patterns. These cluster patterns differed in level, trends and regularity. Depending on the feeding trait, gestation phase and farm, patterns in clusters could mostly differ in average expression of a behaviour, but also could show different increasing, decreasing and changing trends during gestation. Especially feeding traits related to NN-visits showed clusters with more pattern variation, which could reflect behavioural types. Differences in feeding patterns were also observed between farms, gestation phases and sow parities, indicating those factors are important for understanding variation in feeding traits. SOMs proved useful as a tool for first insights. Further analysis to understand the meaning of patterns and to quantify the most relevant aspects of patterns would be valuable in future research.