Improving lameness detection in cows: A machine learning algorithm application.

Area of Science:

• Animal Science
• Agricultural Technology
• Data Science

Background:

• Diverse data-generating technologies offer potential for early disease detection and improved animal well-being in livestock farming.
• Integrating routinely collected farm data with advanced sensor information is crucial for proactive health management.

Purpose of the Study:

• To predict new lameness events in dairy cattle using a combination of farm, herd, weather, and high-frequency sensor data.
• To evaluate the impact of sensor data on the precision and accuracy of lameness prediction models.
• To assess the trade-offs between false positives and false negatives in lameness detection.

Main Methods:

• A Random Forest classifier was employed for lameness prediction.
• The Boruta algorithm was utilized for input feature selection.
• Partial dependence plots were used to assess the effects of individual features.
• Data from 6 dairy farms, including routine, weather, and high-frequency sensor data, were analyzed.

Main Results:

• Precision scores of up to 93% were achieved for predicting lameness up to 3 weeks in advance.
• A balanced accuracy of 79% was obtained when using data from the last 3 weeks.
• Removing sensor data tended to decrease prediction precision, particularly for longer prediction windows.
• A larger dataset without sensor data showed similar balanced accuracy but a significant drop in precision, highlighting the value of sensor data.

Conclusions:

• High-frequency sensor data significantly enhances the precision of lameness prediction models in dairy cattle.
• While sensor data is valuable, high-resolution data from systems like automated milking systems can partially compensate for its absence.
• The findings underscore the importance of sensor integration for improving early disease detection and animal welfare in dairy farming.