Sharing pain: Using pain domain transfer for video recognition of low grade orthopedic pain in horses.
Authors: Broomé Sofia, Ask Katrina, Rashid-Engström Maheen, Haubro Andersen Pia, Kjellström Hedvig
Journal: PloS one
Summary
Early detection of orthopedic disease in horses could prevent unnecessary euthanasia, yet subtle pain behaviours—infrequent, variable, and difficult for even experienced observers to label consistently—remain poorly recognised on video. Broömé and colleagues leveraged domain transfer learning to overcome this annotation challenge, training a convolutional neural network on video data from horses subjected to acute experimental pain (where behavioural responses are clearer and more reliably labelled) and then applying it to detect the far subtler pain signatures of naturally occurring orthopedic conditions. The model trained exclusively on experimental pain performed comparably to expert human observers when identifying orthopedic pain in independent video footage, suggesting that pain-related movement patterns generalise across different aetiologies and severities. These findings have significant implications for developing automated monitoring systems in practice: rather than requiring large datasets of ambiguous, real-world lameness cases, practitioners and researchers can build robust detection algorithms using well-controlled experimental protocols, then deploy them to identify early-stage degenerative disease before clinical signs become obvious to the naked eye. The authors' discussion of best practices for animal behaviour datasets and their commitment to open-source code provision offer a pragmatic framework for integrating computer vision into equine welfare monitoring and diagnostic workflows.
Read the full abstract on PubMed
Practical Takeaways
- •Automated video-based pain recognition systems could help detect early orthopedic disease before clinical signs become obvious, potentially preventing unnecessary euthanasia
- •Current technology requires training on clear acute pain signals before it can reliably detect subtle chronic pain behaviors in individual horses
- •Implementation of such systems in practice would need to account for high variability in how individual horses express low-grade pain
Key Findings
- •Machine learning models trained on acute experimental pain videos can transfer to recognize subtle chronic orthopedic pain in horses
- •Pain recognition from video is challenging due to subtle, sparse, and variable behavioral expressions in orthopedic conditions
- •Domain transfer methods improve detection of low-grade orthopedic pain when training data from acute pain models is applied to chronic pain datasets
- •Expert human labelers show variable agreement on subtle orthopedic pain recognition, establishing baseline difficulty of the task