Machine learning could have role in pain detection in horses – study – Horsetalk

A screen shot of a video of a horse in the study with the visible predicted marker nose (green), withers (red) and tail (blue). Photo: Kil et al. https://doi.org/10.3390/ani10122258

Automated video tracking of stabled horses is a promising tool that, when combined with machine learning, could successfully track pain-related behaviour, according to researchers.

Such a system would be especially useful in a clinical setting, monitoring unwell horses or those recovering from surgery.

Researchers with the University of Veterinary Medicine Vienna set out to evaluate how a video-based automatic tracking tool performed in recognising the activity of stabled horses in a hospital setting.

Nuray Kil, Katrin Ertelt and Ulrike Auer, writing in the journal Animals, said it was well established in veterinary medicine that pain triggers behavioural changes in animals.

Detailed knowledge of both normal and pain-related behaviours in equines is crucial to properly evaluate pain.

Although the presence of strangers or unfamiliar surroundings may mask pain-related changes, even subtle variations may become apparent if behaviour is thoroughly analysed, they said.

In horses, pain is typically scored manually, they said.

Various pain assessment scales, such as the Composite Pain Score and the Horse Grimace Scale, have been developed and proven useful in the assessment of postoperative pain.

However, all methods have limitations and present practical challenges, they noted. For example, horses may be seen only for a short time, and inexperience by the observer may increase the risk of underestimating pain.

A total of 34 horses were used in the study. All were patients of the universitys equine teaching hospital and were housed in box stalls with free access to water, and roughage feed four times a day.

Video recordings were taken using an action camera and a time-lapse mode.

The videos were processed using the convolutional neural network Loopy for automated prediction of three body parts the nose, withers and tail. Development of the model was carried out in several steps.

Ultimately, the body parts were detected with a sensitivity of more than 80% and an error rate between 2% and 7%, depending on the body part. Put simply, the technology was able to identify the pose of the horses with an accuracy and sensitivity of more than 80%.

The results provide a crucial step toward developing algorithms for the automated recognition of behaviour through machine learning.

In the long term, this technology will not only improve the detection of acute and chronic pain in veterinary medicine, but also provide improved and new insights for behavioural research in horses, they said.

The findings will help to develop the automated detection of daily activity, to meet the ultimate objective of objectively assessing the pain and wellbeing of horses.

The study team gave examples of the kinds of insights possible with automated tracking.

For example, the position of a horse in the box in relation to the door can be determined over a longer period of time, or frequent weight-shifting during rest could be detected through nose movement.

The addition of other markers, such as the hooves or ears, would improve the observation of behaviour, they said.

Kil, N.; Ertelt, K.; Auer, U. Development and Validation of an Automated Video Tracking Model forStabledHorses. Animals 2020, 10, 2258.

The study, published under a Creative Commons License, can be read here.

Go here to see the original:
Machine learning could have role in pain detection in horses - study - Horsetalk