Equine Facial Action Coding System for determination of pain-related facial responses in videos of horses

Abstract

1AbstractDuring the last decade, pain scales including facial expressions as indicators of pain have been developed for horses, mostly relying on direct observations or inspection of images. Despite differences in the research conditions and methodology the different scales focus on the same regions of the face, corresponding to moveable facial muscles related to the ears, eyes, nostrils, lips and chin. However, a detailed comparison of the facial activities occurring during pain is not possible. We used a Facial Action Coding System modified for horses to code and analyse video recordings from an earlier study of acute short-term experimental pain and from clinical cases with and without pain. We demonstrated for the first time EquiFACS based changes to pain in video of horses, using traditional statistical methods based on frequency, and novel analyses based on sliding observation windows and co-occurrence of facial actions. The most prominent differences of the experimental horses were related to the lower face actions chin raiser and nostril dilator, while less prominent, but significantly more frequent actions were related to the eye region, inner brow raiser (AU101), increased eye white (AD1), half blink (AU47), and ear rotator (EAD104). Ears forward (EAD101) and eye blink (AU145) were not associated to pain. Based on this we selected the two lower face actions for analysis of the clinical videos, and found that their co-occurrence within a window of 10 to 15 second gave 100% positive predictive values, as compared to the rating from three expert pain raters. Using our developed co-occurrence analyses we were surprised to detect that the chance of identifying three or more of the facial actions related to pain in 0.04 second sequence, corresponding to one frame, was below 3%, indicating that use of randomly selected images for pain scoring may be a very insensitive method.

Improving the Reliability of Scale-Free Image Morphometrics in Applications with Minimally Restrained Livestock Using Projective Geometry and Unsupervised Machine Learning

“…For anatomical subregions with fleshier traits, namely the eye and muzzle metrics, changes in facial expression between photos may partially account for correlations amongst error terms. As contractions of facial muscles cause some length measures to shorten, others must grow longer in compensation, creating systematic error structures that would be consistent across cows by virtue of their shared muscular anatomy [ 6 ]. That changes in the environment between photographs might elicit a response in facial musculature that would produce systematic changes in the annotations of anatomical landmarks in the fleshier regions of the face is not really a surprising result.…”

“…Landmark-based approaches to image analysis are a natural choice for such exploratory analyses. While not informationally complete, they can be easily implemented with manual annotation protocols in a range of programs without the extensive image preprocessing required by embedding techniques [ 5 , 6 ]. How the information contained in coordinate vectors is then compressed into 1D biometrics suitable for statistical analysis varies widely with the use case.…”