Pigs on the move: classifying days with tail lesions using motion sensor data

Continuous activity monitoring using motion sensors provides a non-invasive method to assess pig general behavioural activity and detect potential welfare issues on group level in commercial housing systems. In the current study, motion sensors were installed in rearing and fattening pens on four farms (18 pens in total, 18–40 pigs per pen) over multiple trials. Based on daily tail inspections 4977 observation days were classified into control (C, without tail lesion), pre-tail lesion (P, last seven days before tail lesion) and tail lesion (TL) days. Binary classification ablation studies were used to evaluate and understand model robustness, using pairs of classes and varying time windows for P days. Results indicated that a time window of 7 d before the occurrence of tail lesions showed improved classification performance compared to other time windows, allowing the binary distinction of C, P and TL days. Feature selection revealed that not only motion sensor data but also time of year and housing-related variables influenced model performance. The final multiclass model, based on a 7-day time window for P days, achieved a mean testing accuracy >0.7, ranging from 0.75 to 0.85 for the individual classes and lower accuracies for P days compared to TL and C days. However, leave-one-out cross-validation revealed limited generalisation to unseen farms, likely because the general behavioural activity features are not fully independent of farm. Overall, this study demonstrates the potential of motion sensor data for the early detection of tail lesion while indicating that further research is required to improve robustness and generalisability across farm environments.