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

Area of Science:

• Veterinary Medicine
• Biomechanical Engineering
• Machine Learning

Background:

• Detecting canine lameness and pain is challenging, limiting diagnosis and treatment.
• Clinical gait analysis offers insights but struggles with subclinical or multi-limb lameness.
• Machine learning (ML) presents a potential solution for objective lameness assessment.

Purpose of the Study:

• To develop a proof-of-concept ML model for detecting lameness in dogs.
• To utilize temporospatial and kinetic gait parameters for lameness detection.
• To explore ML's capability in identifying subclinical lameness and localizing gait abnormalities.

Main Methods:

• Collected gait data from 119 normal and 54 injured dogs at a walk, and 98 normal and 31 injured dogs at a trot, using a pressure-sensitive treadmill.
• Trained classification models on 85% of the data and tested on the remaining 15% for each gait style.
• Employed f1 scoring for model training and point biserial correlation for parameter analysis.

Main Results:

• The walking model accurately detected all clinical lameness cases in the test set.
• The walking model did not detect subclinical lameness.
• The trotting model identified one of two subclinical lameness cases.
• Parameter patterns suggest potential for future models to distinguish lameness locations.

Conclusions:

• ML-assisted gait analysis is a viable approach for detecting canine lameness.
• Current models show high accuracy for clinical lameness but require refinement for subclinical cases.
• Future ML models could assist in pinpointing lameness location and type, improving diagnostic strategies.