Pain assessment in horses using automatic facial expression recognition through deep learning-based modeling.
Authors: Lencioni Gabriel Carreira, de Sousa Rafael Vieira, de Souza Sardinha Edson José, Corrêa Rodrigo Romero, Zanella Adroaldo José
Journal: PloS one
Summary
# Editorial Summary: Automated Pain Detection in Horses via Facial Recognition Researchers in Brazil developed a machine learning system to automate pain assessment in horses by training a convolutional neural network to recognise facial expressions based on the Horse Grimace Scale, addressing a key limitation of current practice where skilled observers must be present for extended periods and may inadvertently influence the animal's behaviour. Seven horses were video-monitored from feeders across six days (two pre-castration and four post-operative), with the resulting facial image database used to train the algorithm to classify pain into three levels (absent, moderate, obvious) or two categories (absent/present). The system achieved 75.8% accuracy for three-level classification and 88.3% for binary classification, demonstrating genuine potential for real-time, continuous pain monitoring without human presence bias. Whilst further refinement is needed before routine clinical implementation, automated facial recognition could revolutionise post-operative monitoring, lameness evaluation, and chronic pain management in equine practice, particularly where 24-hour surveillance is otherwise impractical. For farriers and veterinarians, this technology offers a pathway towards objective, observer-independent pain documentation that could improve early intervention outcomes and support more evidence-based treatment decisions.
Read the full abstract on PubMed
Practical Takeaways
- •Automated pain monitoring systems using facial recognition may improve early detection of post-operative complications and guide treatment decisions without handler interference
- •Current accuracy (75.8% for three-level classification) suggests further refinement needed before clinical implementation, but demonstrates feasibility of machine vision for equine pain assessment
- •This technology could reduce labor demands on facilities for pain monitoring post-surgery and enable around-the-clock assessment in high-risk cases
Key Findings
- •Convolutional Neural Network algorithm achieved 75.8% accuracy in classifying pain into three levels (not present, moderately present, obviously present) using facial expressions
- •Binary classification of pain presence vs absence reached 88.3% accuracy using the same machine learning model
- •Automated video-based pain assessment system could enable continuous monitoring without observer presence, reducing behavioral confounding and improving early pain detection