Researchers have employed machine learning in combination with inertial measurement units to accurately classify the various gaits of horses.
“The human eye has thus far served as the ‘gold standard” for gait classification,” Dr Filipe Serra Braganca and his fellow researchers noted in the journal Scientific Reports.
“It is clear from the current study, however, that human visual and subjective assessment is not optimal for this purpose.
“This observation is in line with other studies evaluating human assessment of equine locomotion, mainly in relation to the evaluation of lameness in clinical situations,” the multidisiplinary study team reported.
There, too, human subjective assessment proves less than optimal, as it is affected by the limitations of the human eye and the proneness to bias.
The authors noted that, for centuries, humans have been fascinated by the natural beauty of horses in motion and their different gaits.
Scientific work on gaits in animals was pioneered by Milton Hildebrand. In a ground-breaking article published in Science in 1965, he described a gait classification paradigm. Using manually and subjectively digitized high-speed films, Hildebrand and his colleagues categorized four-legged locomotion into walking and running, and into symmetrical and asymmetrical gaits.
“These relatively simple classification categories have, however, been questioned as to how accurate they are in reliably distinguishing gaits and to what extent they can explain the complex gait patterns generated by the multiple components of the locomotor apparatus of quadrupedal animals.
“More recently, multidimensional approaches have been used, challenging the old dogma.”
The researchers said the development of reliable, automated methods for real-time objective gait classification in horses is warranted.
Four breeds analysed
For their study, they used a full-body network of wireless, high sampling-rate sensors combined with machine learning to fully automatically classify gait.
In all, they used data from 120 horses equipped with seven motion sensors. The work focused on 7576 strides involving eight gaits — the walk, trot, left canter, right canter, tölt, pace, paso fino and trocha.
Four breeds were used: the Colombian Paso, Icelandic, Warmblood and Franche Montagne.
Several machine-learning approaches were used, both from feature-extracted data and from raw sensor data. The best model achieved 97% accuracy, they reported.
“Our technique facilitated accurate, gait classification that enables in-depth biomechanical studies and allows for highly accurate phenotyping of gait for genetic research and breeding.”
The authors said the footfall pattern and the sequence of footfalls can be defined for each gait.
“Some specific features of the gaits can easily be identified, such as symmetry and laterality,” they said. “However, for some gaits such as walk, tölt and paso fino, these variables do not fully discriminate between the gait classes.”
Similarly, the time between strides can be enough to differentiate between some gaits, such as for the walk and trot. “But in other gaits, some of these features overlap, such as stance duration for paso fino and trocha.”
This, they said, highlights the need for multidimensional classification models for the comprehensive classification of all gaits.
The authors said they were able to use technology to extend Hildebrand’s original equine gait paradigm from 1965, showing that reality is more complex and ambiguous, and less straightforward than the original concept.
“This is not unexpected, since Hildebrand’s original model was two-dimensional.
“Our results confirm that gaits are in fact separated by multidimensional planes and that accurate classification can be achieved for this unique diverse gait data using automated approaches that include minimal preprocessing of the signal.”
The researchers say the models used in the study open a new world of possibilities, such as research into the genetics of gait.
“Most equine genetic studies focusing on locomotion, either related to gait or sports performance, require precise phenotyping in order to discriminate between trends in populations or sub-populations.
“Gait phenotyping is still performed subjectively in most of these studies and thus much less accurate than desirable; we therefore believe that our more accurate methods will allow forthcoming studies to understand the genotype–phenotype association of gaits in greater detail.”
They said their approach had the potential for use in other quadrupedal species without the need for developing gait/animal specific algorithms.
Serra Bragança, F.M., Broomé, S., Rhodin, M. et al. Improving gait classification in horses by using inertial measurement unit (IMU) generated data and machine learning. Sci Rep 10, 17785 (2020). https://doi.org/10.1038/s41598-020-73215-9