Behavioural complexity in fattening pigs II. Accelerometer-based validation of a novel welfare indicator with potential for automation

The interest in automated methods of continuously monitoring the behaviour of farm animals, particularly by means of accelerometers, is growing because behaviour is of specific importance for the assessment of welfare beyond health. At the same time, the need for animal-based and individual-specific welfare indicators has been theoretically recognised. Behavioural complexity could be such an indicator with the potential for automation. In this study, we investigated accelerometer data from 97 fattening pigs – recorded on 185 days in three age classes, two seasons, and four housing levels – to assess behavioural complexity. First, we used a binary classifier to automatically predict phases of low and high activity as a basis for calculating behavioural complexity from the accelerometer data. Second, we trained and validated a random forest classifier to predict eight different behavioural states. The binary classification of activity phases achieved an accuracy of 77 % and the random forest model an accuracy of 84 %. For the binary and the random forest data, five and 12 complexity-features, respectively, were reduced using PCA and the resulting PCs were analysed using mixed effects models. In the binary data, general complexity (PC 1) was higher with higher housing levels (p [housing level] = 0.01; p [global] = 0.07) and duration variability (PC 2) tended to be so (p [housing level] = 0.08; p [global] = 0.38). Additionally, there was a tendency towards higher general complexity in older pigs (p [age] = 0.06). For the random forest data, general complexity (PC1) was lower on average with higher housing levels (without statistical support: p [housing level] = 0.14; p [global] = 0.48) and there was a tendency for a higher transition rate (PC2) with higher housing levels (p [housing level] = 0.07; p [global] = 0.52). These results suggest that the higher behavioural complexity seen in more animal-friendly housing environments can also be observed in automatically recorded data classified in binary phases of low and high activity. The classifier-features calculated from accelerometer data collected from head movements could not unequivocally differentiate between the different behaviours. Therefore, the random forest model could not precisely predict the behaviours needed as the basis to calculate behavioural complexity. All observed patterns were relatively weak such that the observed sensitivity of these automatically recorded variables (PCs) in respect to the housing system is unlikely to translate easily into sufficient specificity for future assessment of welfare of individual animals. • Automated behaviour monitoring helps in assessing welfare beyond health. • Behavioural complexity as an individual welfare indicator with automation potential. • Binary and random forest classifiers with fair accuracy in behavior prediction. • Higher behavioural complexity observed in more animal-friendly housing systems. • Weak observed patterns limit specificity for automated welfare assessment.