Machine learning-assisted gait analysis for lameness detection in heterogeneous dog populations

Definitively detecting the presence and location of pain and lameness in a dog can be challenging. Clinical gait analysis can be a useful tool to understand, diagnose, monitor, rehabilitate, and design treatment plans for a variety of musculoskeletal and neurological abnormalities. However its use to detect and distinguish abnormalities can still be limited, particularly in cases of subclinical lameness or multi-limb involvement. This study aimed to develop a proof-of-concept model for machine learning-assisted detection of lameness using temporospatial and kinetic gait parameter measurements. A total of 119 clinically normal dogs and 54 injured dogs were measured using a pressure-sensitive treadmill at the walk, while 98 clinically normal dogs and 31injured dogs were measured at the trot. A classification model was trained on 85% of the data for each gait style using an f1 scoring method and tested with the remaining 15% of the data. The walking classification model successfully identified all clinical lameness cases in the test data, but was not able to detect subclinical lameness cases. The trotting classification model was successful in detecting one of two subclinical lameness cases. Point biserial correlation analysis suggests unique patterns of parameters may enable future models to distinguish lameness locations. These findings support the development and implementation of machine learning methods to detect abnormalities in canine gait, including cases of subclinical lameness, and to assist in determining the location and type of lameness a dog may be experiencing to provide a better understanding of where and how to utilise further diagnostic methods.