Simplify the interpretation of alert lists for clinical mastitis in automatic milking systems.

摘要

Based on sensor measurements, an automatic milking system (AMS) generates mastitis alert lists indicating cows which are likely to have clinical mastitis (CM). Because of the general assumption of equal probabilities of developing CM for all cows, all alerts on the list have the same success rate. As a consequence, it is not possible to rank-order the alerts in terms of their likelihood of CM. In practice, the performance of a CM detection system is not only based on the sensitivity (SN) and specificity (SP) of the system, but is also influenced by the prior probability of a cow having CM. This study illustrates the idea of using cow-specific prior probabilities of CM, based on non-AMS information, to provide a rank-order on the alerts from an AMS. A tree-augmented naive Bayesian network was trained from available data to determine these cow-specific prior probabilities for CM. The graphical structure of the network and the probability tables for its variables in the network were based on data from 274 Dutch dairy herds that recorded each case of CM over an 18-month period. The final data set contained information on a total of 5363 CM cases derived from 28,137 lactations and 22,860 cows. The available prior cow information (parity, days in milk, season of the year, somatic cell count history and CM history) was included as variables in the network. By combining the cow-specific prior probabilities of CM with the SN and SP of the detection system of the AMS, the computed success rates can be used to discriminate between CM alerts. Our illustrations indicate that the success rate might range from 3 to 84%, while assuming an equal overall probability would result in a success rate of 21%. Using the computed success rates, the CM alerts on an alert list can be rank-ordered, thereby providing the dairy farmer information about which cows have the highest priority for visual inspection for CM