Quantification of pain from facial expression characteristics assessed from photos of horses and donkeys

Abstract

In the past decade, there has been more attention paid to the evaluation of pain in animals. Assessment of pain has become more important to horse and donkey welfare. This study describes the assessment of objective scoring on photos of horses and donkeys measured with two newly developed pain scales: the Horse Pain Face (HPF) and the Donkey Pain Face (DPF). In this research, 1654 horse photos and 534 donkey photos were used. The photos were scored by two trained observers, with variable levels of equine experience, to determine the inter- and intra-observer reliability by intra-class correlation analysis (ICC), determination of Cronbach’s alpha and Bland Altman analysis. For the intra-observer reliability 20% of the photos from both the horse and donkey photo sets were rescored (n=331 ; n=107 respectively). The assessment of photos were performed simultaneously and the observers did not discuss their findings. This research also analyzed the HPF and DPF inter-observer reliability of three observers. The third observer scored 20% photos that were used to asses intra-observer reliability. The HPF scored a very good inter-observer reliability (Cronbach’s alpha = 0.92 ; P<0.001) and the DPF scored an acceptable inter-observer reliability (Cronbach’s alpha = 0.79 ; P<0.001), while good intra-observer reliabilities were found for both the HPF and the DPF by the first two observers. (Observer 1: Cronbach’s alpha = 0.81 for the HPF ; Cronbach’s alpha = 0.84 for the DPF ; P<0.001) (Observer 2 : Cronbach’s alpha = 0.88 for the HPF ; Cronbach’s alpha = 0.97 for the DPF ; P<0.001). Both the HPF and the DPF scored an excellent inter-observer reliability by the three observers. (Cronbach’s alpha = 0.91 for the HPF ; P<0.001 (n=331) and Cronbach’s alpha = 0.92 for the DPF; P<0.001 (n=107)). Both pain scales scores proved valid for objective pain and repeatable assessment of pain in horse and donkey photos. This study showed the importance of training to obtain valid and reproducible results with the two pain scales. The results of this study will be used in a future project using neural networks and deep learning techniques for automated pain scoring from facial expressions.