Optimising the selection of welfare indicators in farm animals

Risk assessment (RA) frameworks are increasingly being applied to improve the welfare of farmed animals. These frameworks have at their core, a logic chain linking welfare hazards (risks) with one or more welfare consequences which, in turn, are each measured by one or more welfare indicators. Effective and efficient monitoring of animal welfare often involves the selection of a subset of indicators from a large pool. Selecting 'iceberg indicators' could be advantageous due to their association with multiple welfare consequences. However, no standardised, data-driven method exists to select optimal combinations under practical constraints. This study addresses this gap by creating an algorithmic approach to optimise indicator selection.

The work was conducted in six phases: (1) construction of a structured database of welfare indicators; (2) a proof-of-concept study; (3) design of a greedy selection algorithm; (4) enhancement of the algorithm using branch-and-bound and backtracking methods; (5) performance and sensitivity testing, and (6) creation of two case studies. A dataset of 382 animal welfare indicators across seven farm species was compiled from scientific opinions published by the European Food Safety Authority (EFSA) and from other published literature. The EFSA scientific opinions contain data acquired through a rigorous process of literature reviews and expert elicitation and consensus panels to link welfare indicators with their associated welfare hazards and welfare consequences. To enable algorithm development, the Coverage of each welfare indicator was first determined by calculating the number of unique welfare consequences to which it was linked. Metadata such as the Impact of welfare consequence [Low (1) or High (2)], Ease of hazard mitigation [Easy (1), Moderate (2) or Difficult (3)], and Ease of indicator use [Easy (1), Moderate (2) or Difficult (3)] was generated through an expert elicitation process. These data were standardised using max-min normalisation across all criteria, and an objective function was defined which enabled indicator subset selection according to various user-defined criteria. Optimisation was performed using both a greedy algorithm and an enhanced algorithm incorporating backtracking and branch-and-bound solvers. Algorithm performance and robustness were evaluated through sensitivity analyses, scenario testing, and computational benchmarking.

The greedy algorithm offered computational efficiency but incorporated suboptimal plateaus in Coverage as additional indicators were combined. The enhanced algorithm identified globally optimal combinations within 0.2 s for all species, regardless of problem size. In a broiler chicken case study, the enhanced algorithm removed indicators that were moderately difficult to use. A pig case study showed that the enhanced algorithm combined the same welfare indicators as the greedy algorithm but validated the added value of multi-criteria scoring by identifying high-impact, easy-to-implement indicators suitable for welfare certification.

The enhanced algorithm was able to move beyond the selection of iceberg indicators, by incorporating multiple selection criteria to inform welfare indicator choice. The enhanced algorithm is data-agnostic and enables users to optimise indicator selection with diverse datasets spanning research, industry, and policy contexts. Its flexibility supports the development of tailored applications for different stakeholders. Future work should explore processes to determine weighting values, scenario testing, robustness, and stakeholder engagement to maximise both relevance and practicality.