Related Experiment Video
Updated: May 6, 2026
08:20
Author Spotlight: AI-Driven Trypanosome Species Detection from Microscopic Images
Published on: October 27, 2023
2.9K
Detection of clinical mastitis with sensor data from automatic milking systems is improved by using decision-tree
C Kamphuis1, H Mollenhorst, J A P Heesterbeek
1Department of Farm Animal Health, Faculty of Veterinary Medicine, Utrecht University, Utrecht, the Netherlands. C.Kamphuis@uu.nl
Journal of Dairy Science
|July 27, 2010
Summary
A new decision-tree model for automatic milking systems significantly reduces false mastitis alerts. This clinical mastitis (CM) detection tool improves specificity while maintaining sensitivity for early disease detection in dairy cows.
Area of Science:
- Veterinary Medicine
- Animal Science
- Dairy Science
Background:
- Clinical mastitis (CM) is a significant concern in dairy farming, impacting animal welfare and milk production.
- Automatic milking systems (AMS) offer potential for early CM detection through sensor data, but require highly sensitive and specific models.
- Current CM detection models in AMS often generate false alerts, necessitating improved diagnostic tools.
Purpose of the Study:
- To develop and validate a decision-tree induction model for detecting clinical mastitis (CM) in dairy cows milked with AMS.
- To achieve high sensitivity (Se), particularly for severe CM cases, and very high specificity (Sp).
- To enable CM alerts at the earliest stage of infection during quarter milking (QM).
Main Methods:
- Collected sensor data (electrical conductivity, color, yield) and farmer-recorded visual CM observations from 9 Dutch dairy herds over 2.5 years.
- Integrated sensor data within a 24-h window around visual CM assessments for 3.5 million quarter milkings (QM).
- Trained a decision-tree model on healthy and diseased QM data, then validated it on a separate test set, comparing performance against existing AMS models.
Main Results:
- The decision-tree model reduced false-positive alerts by over 50% compared to current AMS models, while maintaining comparable Se.
- At 99% Sp, the model detected 40% of CM cases, including 64% of severe CM cases, but only 12.5% of watery milk cases.
- Increasing the time window for data integration improved Se from 40% to 66.7%, though 100% detection remained unachievable with sensor data alone.
- Model performance varied significantly across herds, indicating challenges in developing a universally applicable generic CM detection model.
Conclusions:
- The developed decision-tree model offers improved specificity for CM detection in AMS, reducing false alerts.
- While effective for severe cases, the model's sensitivity for early or mild CM, like watery milk, requires further improvement.
- Herd-specific variations in sensor data necessitate further research for a robust, generic CM detection model applicable across diverse dairy farming environments.
