A new AI framework can detect signs of pain in horses by analyzing video, even as the animal moves or turns its head. The system, called SHIC-XE, also generates anatomically consistent explanations for its decisions - a step toward building trust in AI tools for clinical environments where patients cannot speak for themselves.
The research was led by Dr. Marcelo Feighelstein, head of the Artificial Intelligence Systems Engineering Program at Tel Hai University. The work, published in the International Journal of Computer Vision, involved collaborators from the University of Haifa, the Technion, and several international universities. The team solved a persistent problem in explainable AI: traditional 2D saliency maps become unstable and flicker when applied to video, making them unreliable for medical or veterinary use.
Projecting attention onto a 3D model
Standard methods for visualizing what a model focuses on - heat maps overlaid on images - break down when both the subject and the camera move. Frame-to-frame inconsistencies prevent meaningful temporal aggregation. For horses, the challenge is compounded. They are prey animals known to mask pain signals when humans are nearby.
SHIC-XE addresses this by projecting the model's attention onto a fixed 3D representation of the horse's face. This creates a stable anatomical map regardless of head position or camera angle. Feighelstein said the model focuses on biologically relevant regions when making its decision. "We proved that the model not only reaches the correct conclusion, but also focuses on the anatomically relevant regions, especially the ears and cheek muscles, when making that decision," he said.
Beyond veterinary medicine
While the current application targets equine pain, the underlying approach has wider clinical potential. Feighelstein pointed to human patients who cannot communicate: newborns, sedated ICU patients, and people with dementia, stroke, or aphasia. The same technique could analyze neurological movement disorders or assist with surgical video analysis.
The research is part of a broader effort by Feighelstein's lab to build AI video analysis tools for animal welfare. Previous work has covered cats, dogs, rabbits, sheep, and cattle. Dogs present a particular challenge due to extreme variation in facial structure across breeds, from short-nosed Shih-Tzus to long-nosed collies and greyhounds.
Building a bridge between humans and animals
Feighelstein described the core motivation plainly: "We want to give a technological voice to those who have no words, allowing caregivers and professionals to better understand their condition, their emotions, and their suffering." He emphasized that the technology is designed to support veterinarians, not replace them. Continuous monitoring could alert staff when an animal's condition changes during transport, on farms, or in hospital bays when no one is watching.
The current system detects the presence of pain. Future work aims to estimate severity and potentially identify the source - whether inflammation, orthopedic injury, surgical recovery, or another cause. An application based on the model is under development.
Why this matters for IT and development professionals
SHIC-XE demonstrates a concrete solution to the explainability-stability tradeoff in video-based AI. For developers building computer vision systems in regulated industries - healthcare, veterinary medicine, or any domain requiring auditable decisions - the 2D-to-3D projection approach offers a template for producing explanations that hold up under expert review. The framework's quantitative comparison with expert assessments moves explainable AI from a qualitative checkbox toward a measurable engineering requirement.
Your membership also unlocks:
