AI System Reads Pain in Horses—and Shows Exactly Where It Sees Suffering
A new artificial intelligence framework can detect signs of pain in horses from video footage while showing veterinarians precisely which anatomical features influenced its decision. Called SHIC-XE, the system was developed by an international research team led by Dr. Marcelo Feighelstein, Head of the Artificial Intelligence Systems Engineering Program at Tel-Hai University of Kiryat Shmona in the Galilee. Unlike many AI models that produce a prediction without revealing how they reached it, SHIC-XE is designed to provide explanations that remain stable as a horse moves, turns its head, or is recorded from a different viewpoint. The researchers say this combination of recognition and interpretability could help transform AI from a promising experimental technology into a dependable tool for veterinary care and, eventually, human medicine.
Pain is difficult to measure in any patient who cannot describe it verbally. In horses, discomfort may appear through subtle changes in facial expression, ear position, muscle tension, posture, and movement. These signals can be difficult to interpret consistently, even for experienced veterinarians, particularly when pain is mild, fluctuating, or accompanied by fear and stress. The research team trained and evaluated an AI system that analyzes video rather than relying on a single photograph. Its objective is not merely to classify whether a horse appears to be in pain, but also to identify the visual evidence supporting that conclusion. This distinction is important because an accurate prediction can still be clinically unsafe if the system bases its decision on irrelevant details, such as background objects, camera position, lighting, or equipment visible in the scene.
The central innovation of SHIC-XE is a method for producing viewpoint-invariant explanations. Conventional visual explanation techniques often generate heat maps that highlight the parts of an image most associated with an AI prediction. However, those highlighted regions may shift dramatically from one video frame to the next. A horse’s face can change shape within the image as it turns, while the same anatomical feature may appear in different pixel locations because of movement or camera angle. SHIC-XE addresses this problem through dense two-dimensional-to-three-dimensional correspondences. In practical terms, the system maps visual information from each frame onto a fixed three-dimensional representation of the horse’s face, allowing the model’s attention to be compared in anatomical rather than purely image-based coordinates.
That mapping creates a common reference system for the AI’s explanations. Instead of saying that a few changing pixels near the edge of an image influenced its decision, the framework can indicate that attention was consistently directed toward recognizable regions of the horse’s anatomy. This makes the explanation easier to interpret and evaluate. The approach is especially relevant to pain recognition because veterinary assessment often depends on small, localized changes. The position of the ears, tension around the eyes, and alterations in the cheek muscles may carry more diagnostic meaning than the overall appearance of the animal. By projecting the AI’s attention onto a stable facial model, SHIC-XE helps researchers determine whether the system is responding to those meaningful signals or to accidental correlations in the training data.
The researchers tested the framework on three independent datasets representing different clinical conditions: pain following surgery, inflammatory and orthopedic pain, and acute mechanical pain. Across these scenarios, the system achieved F1 scores between 0.67 and 0.80. The F1 score combines precision, which measures how many positive predictions are correct, with recall, which measures how many true cases are detected. A higher score indicates a better balance between identifying animals in pain and avoiding false alarms. Although performance varied across the datasets, the results suggest that the model can recognize pain across multiple clinical contexts rather than learning a narrow visual pattern associated with only one type of injury or treatment.
The most significant result, however, involved the relationship between the AI’s explanations and expert judgment. Veterinary specialists assessed the same cases using the Horse Grimace Scale, a standardized system based on facial indicators of pain. The regions emphasized by SHIC-XE showed statistically significant agreement with the areas considered relevant by the experts, particularly around the ears and cheek muscles. This provides a quantitative way to test not only whether an AI model reaches the correct answer, but whether it reaches that answer for scientifically and clinically credible reasons. Dr. Feighelstein described the result as the first quantitative validation of its kind, demonstrating that the model can focus on anatomically relevant regions while making its prediction.
The ability to inspect an AI system’s reasoning is becoming increasingly important as machine learning moves into healthcare and animal welfare. Many high-performing neural networks remain difficult to interpret because they learn complex relationships distributed across millions of parameters. An explanation generated after the prediction is not automatically proof that the model used valid evidence. It must be stable, anatomically meaningful, and capable of being compared with independent expert assessments. SHIC-XE was designed around those requirements. Its anatomical coordinate system allows explanations to be aggregated across videos and subjects, making it possible to study whether the model repeatedly attends to the same regions when similar signs of pain are present.
The project is part of a broader effort led by Feighelstein and his collaborators to use AI as a bridge between humans and animals. Earlier work by the research community has explored automated recognition of pain and emotion in cats, dogs, rabbits, sheep, and cattle. Each species presents different anatomical structures, behavioral signals, and sources of variation, meaning that models cannot simply be transferred without careful adaptation. Horses are particularly important in veterinary research because their pain can be associated with surgery, orthopedic disease, inflammation, and acute injury, while their size and behavior can make direct monitoring challenging. A video-based system could eventually support continuous observation and alert caregivers to changes that might otherwise be missed.
The implications extend beyond equine medicine. A reliable method for understanding what an AI system sees in moving images could be valuable wherever patients cannot communicate clearly. Potential applications include pain assessment in newborns, monitoring sedated or mechanically ventilated intensive-care patients, supporting people with dementia, and analyzing neurological movement disorders. Similar techniques might also assist surgeons and clinicians by making video-based decision systems more transparent. In each case, the underlying challenge is the same: an algorithm must provide more than a label. It must offer evidence that clinicians can inspect, question, and compare with established medical knowledge.
The study was led by Tel-Hai University of Kiryat Shmona in the Galilee in collaboration with Prof. Anna Zamansky and the Tech4Animals Lab, Prof. Ilan Shimshoni of the University of Haifa, students Omer Bibi and Ofer Rosenbaum from the Technion, and researchers from the University of Bern, the University of Milan, the University of São Paulo, and Newcastle University. Published in the International Journal of Computer Vision, the work presents SHIC-XE as both a tool for equine pain recognition and a framework for building more accountable visual AI. By linking predictions to stable three-dimensional anatomy, the researchers aim to open the black box of machine learning—and give caregivers a clearer, evidence-based way to understand when an animal may be suffering.
Subject of Research: Animals
Article Title: SHIC-XE: Viewpoint-Invariant Explainability via Dense 2D-3D Correspondences: an Application to Equine Pain Recognition
News Publication Date: 10-Jul-2026
Web References: https://link.springer.com/article/10.1007/s11263-026-02910-3; https://www.telhai.ac.il/en
References: International Journal of Computer Vision, DOI: 10.1007/s11263-026-02910-3
Image Credits: Photo courtesy of Tel-Hai University of Kiryat Shmona in the Galilee
Keywords
Artificial intelligence, explainable AI, equine pain recognition, veterinary medicine, horse welfare, computer vision, machine learning, facial expression analysis, Horse Grimace Scale, medical AI, animal behavior, three-dimensional modeling
