 |
| © Lifeontheside Dreamstime.com |
As a prey species, horses instinctively mask signs of discomfort to avoid appearing vulnerable, meaning even experienced clinicians can struggle to identify pain in its early stages. New research from an international team of scientists has developed an artificial intelligence (AI) system that could eventually provide an additional tool to help monitor equine welfare.
The researchers, led by Dr Marcelo Feighelstein of Tel Hai University, Israel, have developed a new AI framework called SHIC-XE (Stable Heatmap Integration and Comparison for Explainable AI). Unlike previous AI systems, SHIC-XE not only detects signs that may indicate pain in horses but also provides a clear explanation of how it reached its conclusions.
Current AI systems often use coloured "heat maps" to show which parts of an image influenced a decision. While these can work reasonably well with still photographs, they become unreliable when analysing video. As the horse or camera moves, the highlighted areas can flicker from one frame to the next, making it difficult to understand exactly what the AI is using to identify pain.
SHIC-XE overcomes this limitation by mapping video images onto a stable three-dimensional model of the horse's face. This allows the system to keep its focus on the same anatomical structures, even when the horse changes head position or is viewed from different angles. Rather than producing inconsistent heat maps, the AI consistently relates its decisions to specific facial regions.
Importantly, the researchers showed that the AI concentrated on areas that experienced veterinarians also regard as important indicators of discomfort, particularly the ears and cheek muscles. By comparing the AI's attention with scores from a validated equine pain scale, they demonstrated meaningful agreement between the model and expert assessments.
This represents one of the first studies to provide a quantitative measure of whether an AI explanation matches established veterinary knowledge, rather than relying solely on subjective visual interpretation. The work also highlights that previous claims about the clinical usefulness of AI explanations may need further validation if they were based only on visual inspection.
The system was evaluated using three separate datasets containing videos of horses showing different levels of pain. Its performance was measured using an F1 score, which combines two important measures: how often the system correctly identified horses in pain and how often it avoided incorrectly classifying horses as being in pain. An F1 score of 1.00 would represent perfect performance, while a score of 0 would indicate very poor performance. The system achieved scores of 0.67, 0.80 and 0.70 across the three datasets, suggesting that it performed reasonably well, although it was not perfect.
The researchers used leave-one-subject-out validation, meaning that the system was tested on horses whose videos had not been used to train it. This provides a stronger indication of whether the AI can work with unfamiliar horses rather than simply memorising the animals in its training data.
The researchers stress that the technology is intended to support, rather than replace, veterinary judgement. In the future, AI could provide continuous monitoring in stables, veterinary hospitals or during transport, alerting staff when a horse's behaviour suggests developing pain, even when no one is actively observing it.
Although the current system simply identifies the presence of pain, the research team hopes future versions may be able to estimate pain severity and perhaps distinguish between different causes, such as orthopaedic injury, inflammation or post-operative discomfort. However, these capabilities remain future goals and have not yet been demonstrated.
While SHIC-XE was developed using horses, the underlying approach could have much wider applications. Similar techniques may eventually assist with pain assessment in other animal species and even in human patients who cannot communicate verbally, such as newborn babies or sedated intensive care patients. For equine medicine, however, the research represents an important step towards AI systems that are not only accurate, but also transparent enough for veterinarians to understand and trust.
For more details, see:
Marcelo Feighelstein, Omer Bibi, Ofer Rozenbaum, Nathali Adrielli Agassi De Sales, Guilherme Camargo Ferraz, Ilan Shimshoni, Dirk van der Linden, Emanuela Dalla Costa, Annika Bremhorst, Claudia Spadavecchia & Anna Zamansky
SHIC-XE: Viewpoint-Invariant Explainability via Dense 2D-3D Correspondences: an Application to Equine Pain Recognition.
Int J Comput Vis (2026) 134, 342
No comments:
Post a Comment