Exploring equine behavior: Wearable sensors data and explainable AI for enhanced classification.
Authors: Cetintav Bekir, Yalcin Ahmet
Journal: Journal of equine veterinary science
Summary
# Editorial Summary Wearable collar-mounted sensors combining accelerometers, gyroscopes, and magnetometers offer promising real-time insight into equine behaviour, with Cetintav and Yalcin's recent analysis of 18 horses demonstrating 82.3% classification accuracy across 17 distinct behavioural states using a Random Forest algorithm. Beyond raw accuracy, the researchers applied SHAP (Shapley Additive Explanations) analysis to identify precisely which sensor data types drive classification decisions—a crucial transparency step often missing from equine AI applications—revealing that high-intensity locomotion like galloping relies predominantly on accelerometer signals, static postures like standing depend on magnetometer orientation data, and stress indicators such as head-shaking are characterised by rapid gyroscopic angular velocity changes. For equine professionals, these findings translate into practical applications: continuous remote monitoring can flag stress-related or health-concerning behavioural changes without requiring direct observation, whilst the interpretable feature attribution helps validate that the system is identifying clinically meaningful patterns rather than statistical artifacts. The open-source dataset and transparent AI methodology establish a reproducible benchmark that could support wider adoption of behaviour monitoring in veterinary practice, training optimisation, and welfare assessment protocols.
Read the full abstract on PubMed
Practical Takeaways
- •Wearable sensor systems with machine learning can provide objective, real-time monitoring of horse behavior and early stress detection to support veterinary decision-making and welfare management
- •Different sensor types capture different behavioral information—accelerometers for movement intensity, magnetometers for body orientation, and gyroscopes for head/neck dynamics—enabling targeted monitoring based on clinical concerns
- •Explainable AI approaches make these 'black-box' models transparent and actionable, allowing practitioners to understand which specific movement patterns trigger behavior classifications
Key Findings
- •Random Forest model achieved 82.3% accuracy in classifying 17 equine behavior classes using wearable sensor data
- •SHAP analysis revealed accelerometer features dominate locomotion behaviors (galloping), magnetometer data characterize stationary behaviors (standing), and gyroscopic data identify stress-related behaviors (head-shaking)
- •Integration of XAI techniques improved interpretability of machine learning models for equine behavior classification and real-time monitoring applications