Non-Invasive Foaling Prediction in Horses Using Pose Estimation and Machine Learning
Authors: Koyo Uratani, N. P. Martono, Takashi Hatazoe, Hayato Ohwada
Journal: 2025 15th International Conference on Information & Communication Technology and System (ICTS)
Summary
# Editorial Summary: Non-Invasive Foaling Prediction Using Pose Estimation and Machine Learning Predicting parturition accurately remains a significant challenge for breeders, as current methods either require invasive hardware (tail accelerometers or implanted thermistors) with limited warning time or demand intensive round-the-clock monitoring. Researchers at Japanese breeding facilities developed a camera-based system using deep-learning pose estimation to identify postural and behavioural changes associated with approaching foaling, extracting 192 features from skeletal position, limb angles and turning patterns across 19 documented foaling events and over 1,300 hours of negative footage. The best-performing model (trained on LightGBM, XGBoost and bidirectional LSTM algorithms) achieved human-level sensitivity whilst reducing false alarms to 0.054 per hour compared to 0.133 for human observers—and crucially, predicted foaling up to two hours before parturition occurred. For commercial operations, this translates to reducing manual stall checks by 97.8%, meaningfully easing labour demands during busy breeding seasons whilst improving welfare outcomes through earlier notification of complications. The non-invasive approach using standard visible-light cameras also removes the logistical burden and cost of wearable devices, making the technology readily scalable across different farm environments.
Read the full abstract on the publisher's site
Practical Takeaways
- •A non-invasive camera-based system can provide 2-hour advance warning of foaling without requiring tail-mounted devices or chips, reducing monitoring burden on breeding farms
- •Pose estimation technology achieves lower false alarm rates than human observers, potentially reducing unnecessary nighttime stall checks from 24+ to less than 1 per night
- •Implementation requires only standard camera equipment and software, making it cost-effective for commercial operations compared to existing invasive instrumentation
Key Findings
- •Non-invasive pose estimation using standard camera footage successfully predicted foaling up to 2 hours before the event with recall matching human observers
- •Machine learning model reduced false alarms to 0.054 per hour compared to 0.133 per hour for human observers
- •System reduced manual stall checks by 97.8% while maintaining detection accuracy
- •192-dimensional features extracted from posture, limb angles and turning behavior proved effective for foaling prediction