Comparison of machine learning and validation methods for high-dimensional accelerometer data to detect foot lesions...

Area of Science:

• Animal Science
• Veterinary Medicine
• Data Science

Background:

• Lameness is a significant issue in dairy cattle, impacting animal welfare, farm economics, and sustainability.
• Early detection of lameness is crucial for timely treatment and mitigating production losses.
• Automated detection systems using accelerometers show promise for early lameness identification.

Purpose of the Study:

• To evaluate machine learning (ML) methods for detecting foot lesions in dairy cows using accelerometer data.
• To assess the effectiveness of dimensionality reduction techniques and cross-validation strategies for analyzing wide, high-dimensional accelerometer data.
• To provide practical insights for the dairy industry on improving ML model performance for lameness detection.

Main Methods:

• Utilized accelerometer data from 383 dairy cows across 11 herds (20,000 recordings).
• Applied dimensionality reduction techniques, including principal component analysis (PCA) and functional principal component analysis (fPCA).
• Employed cross-validation strategies, emphasizing a by-farm approach for robust model evaluation.

Main Results:

• Dimensionality reduction effectively retains key information from wide accelerometer data, enabling broader ML applications.
• Combining dimensionality reduction with cross-validation significantly improves ML model performance for lameness detection.
• A by-farm cross-validation approach provides a more realistic estimate of general model performance.

Conclusions:

• Dimensionality reduction and cross-validation are essential for effective ML application to high-dimensional accelerometer data in dairy cattle.
• These methods enhance the accuracy of lameness detection, benefiting animal welfare and farm economics.
• The study underscores the importance of independent farm data for validating model generalizability.