Improved and interpretable accelerometer-based farrowing prediction

Predicting the onset of farrowing in sows is critical for improving animal welfare and optimising farm management. Methods driven by explainable artificial intelligence for detecting nest-building behaviour and predicting time to farrowing using accelerometer data from ear tags are presented. These methods are evaluated on a dataset containing farm management data and accelerometer data of 179 sows. During data collection the animals were kept in three different pen types with the possibility of temporary crating. By combining acceleration metrics with prepartum examinations and farm management data, a two-stage model was developed that first detects the onset of nest-building and subsequently predicted the remaining time until farrowing. Various methods, including cumulative sum (CUSUM), Bayesian estimation of abrupt change, seasonality, and trend (BEAST), and a custom model (NestDetect), were compared for nest-building detection, while symbolic regression and deep learning were used to predict farrowing time. For 82.6 % of the sows, it was possible to detect the start of nest-building behaviour in a 48-h window before the onset of farrowing. When nest-building was detected correctly, symbolic regression was able to predict the remaining time to farrowing with a mean absolute error of 9.4 h and delivered interpretable results, while NNs achieved a mean absolute error of 9.6 h without being inherently interpretable. This work emphasises the importance of model interpretability and explainability in precision livestock farming, highlighting that transparent models can facilitate timely, data-driven interventions, while having the same prediction power as non-interpretable models.

Technische Universität Wien
FH Oberösterreich Hagenberg