Towards Machine Recognition of Facial Expressions of Pain in Horses.
Authors: Andersen Pia Haubro, Broomé Sofia, Rashid Maheen, Lundblad Johan, Ask Katrina, Li Zhenghong, Hernlund Elin, Rhodin Marie, Kjellström Hedvig
Journal: Animals : an open access journal from MDPI
Summary
# Editorial Summary: Machine Recognition of Equine Facial Pain Expressions Automating pain recognition in horses could revolutionise early detection and welfare assessment, yet developing reliable systems has lagged far behind similar technology in humans, primarily due to the absence of large annotated datasets and the challenge of validating pain states in non-verbal animals. Researchers at this Swedish institution tackled these barriers using two complementary machine learning approaches: a manual Facial Action Coding System (FACS) combined with automated feature extraction to identify equine facial landmarks and action units, and a recurrent neural network (RNN) method analysing video dynamics rather than individual frames. Results from the FACS approach showed promising automated recognition of specific facial action units from still images, whilst the RNN model demonstrated superior performance in classifying experimentally-induced pain compared to human raters in a small validation cohort, particularly when temporal changes in facial movements were incorporated. The findings suggest that dynamic facial analysis—capturing how expressions change over time—is more informative for equine pain detection than static images alone, indicating a clear pathway toward objective, technology-assisted pain assessment tools for clinical and research applications. For practitioners, this represents meaningful progress toward reducing subjective interpretation in pain scoring and potentially enabling earlier intervention in cases of acute or chronic discomfort.
Read the full abstract on PubMed
Practical Takeaways
- •Automated facial expression analysis tools may eventually provide objective, non-verbal pain assessment in horses—potentially improving welfare monitoring and treatment decisions
- •Current technology shows promise but requires substantial video data collection and validation before clinical implementation
- •Manual facial coding systems combined with machine learning offer an intermediate approach where practitioners could use AI-assisted labeling tools to accelerate pain assessment protocols
Key Findings
- •Automated facial action unit recognition in horses shows promising results using machine learning applied to manually coded Facial Action Coding System data
- •Recurrent neural network end-to-end learning models can classify experimental pain in horses better than human raters when trained on video data with temporal dynamics
- •Major barriers to equine pain recognition include lack of large annotated databases and difficulties obtaining ground truth labels due to horses being non-verbal