Decision support for mastitis on farms with an automatic milking system

Abstract

For an optimal mastitis management on farms with an automatic milking system (AMS), two individual cow decisions are important. First, there is a need for decision support on which mastitis alerts have the highest priority for visual checking for clinical mastitis (CM). In essence, all cows with mastitis alerts have to be checked visually for CM. Because of the suboptimal detection performance of current CM detection models, and therefore the large number of false-positive alerts for CM, in practice farmers do not check all mastitis alerts. Non-AMS information about the cow, such as somatic cell count (SCC) history influences the probability of having CM. It is expected that adding non-AMS cow information can be used to support decisions on which mastitis alerts have the highest priority for visual checking. The second important decision that needs support is the choice of treatment for detected CM cases. Different antimicrobial treatments are available for CM on Dutch dairy farms, differing in antimicrobial compound, route of application, duration, probability of cure and costs. Most CM cases receive a standard intramammary treatment. Additional information sources are hardly used to differentiate in choice of treatment for different cows with CM. Several cow factors (e.g., SCC history) and information from the visual inspection influence the success of treatment. These information sources can be used to support CM treatment decisions. The objectives of the research were (i) to improve automated CM detection by combining sensor and non-AMS cow information and (ii) to improve CM treatment decisions. Based on data of a single research farm it was investigated whether non-AMS cow information (parity, stage of lactation, season, SCC history and CM history) can be used to make a selection of alerted cows that need further investigation for CM. The results indicate that adding non-AMS cow information could not be used to make a distinction between true-positive and false-positive alerts. In addition, data of 9 commercial farms was used to determine whether non-AMS cow information can be used to improve CM detection. CM detection performance of a CM detection model based on sensor measurements was compared to a CM detection model based on sensor measurements and added non-AMS cow information. Results show that if sensor measurements are available, adding non-AMS cow information had no additional value for CM detection. Having an indication about the Gram-status of the causal pathogen of CM can improve CM treatment decisions. Cow information and information from the visual inspection was used to give probability distributions for the Gram-status of CM cases. We found that the greater the probability for Gram-positive or Gram-negative pathogens, the more accurate the classification for the Gram-status. To improve CM treatment decisions it was also investigated whether using different treatments for different cows is economically beneficial. A model was developed to simulate CM cases and the consequences of using different treatments. The results show that effectiveness of different antimicrobial treatments does vary between cows, but differentiation of treatments for different CM cases does not provide economic benefits under current Dutch circumstances.