Development of a new clinical mastitis detection method for automatic milking systems.

This study investigated the potential for accurate detection of clinical mastitis (CM) in an automatic milking system (AMS) using electronic data from the support software. Data from cows were used to develop the model, which was then tested on 2 independent data sets, 1 with 311 cows (same farm but from a different year) and 1 with 568 cows (from a different farm). In addition, the model was used to test how well it could predict CM 1 to 3 d before actual clinical diagnosis. Logistic mixed models were used for the analysis. Twelve measurements were included in the initial model before a backward elimination, which resulted in the following 6 measurements being included in the final model: quarter-level milk yield (MY; kg), electrical conductivity (EC; mS/cm), average milk flow rate (MF; kg/min), occurrence of incompletely milked quarters in each milking session (IM; yes or no), MY per hour (MYH; kg/h), and EC per hour (ECH; mS/cm/h) between successive milking sessions. The other 6 measurements tested but not included in the final model were peak milk flow rate (kg/min), kick-offs (yes or no) in each milking session, lactation number, days in milk (d), blood in milk (yes or no), and a calculated mastitis detection index used by DeLaval (DelPro software; DeLaval International AB, Tumba, Sweden). All measurements were assessed to determine their ability to detect CM as both individual variables and combinations of the 12 above-mentioned variables. These were assessed by producing a receiver operating characteristic curve and calculating the area under the curve (AUC) for each model. Overall, 9 measurements (i.e., EC, ECH, MY, MYH, MF, IM, peak flow rate, lactation number, and mastitis detection index) had significant mastitis detection ability as separate predictors. The best mastitis prediction was possible by incorporating 6 measurements (i.e., EC, ECH, MY, MYH, MF, and IM) as well as the random cow and quarter effects in the model, resulting in 90% sensitivity and 91% specificity with excellent AUC (0.96). Assessment of the model was found to produce robust results (AUC >0.9) in different data sets and could detect CM with reductions in sensitivity and specificity with increasing days before actual diagnosis. This study demonstrated that improved mastitis status prediction can be achieved by using multiple measurements, and new indexes based on that are expected to result in improved accuracy of mastitis alerts, thereby improving the detection ability and utility on farm.