Estimating degree of mastitis from time-series measurements in milk: a test of a model based on lactate dehydrogenase measurements.

The aim of this study was to test a model for mastitis detection using a logic that allows examination of time-related changes and a progressive scale of mastitis state (i.e., not using specificity/sensitivity). The model produces a mastitis risk (MR) for individual cows on a scale from 0 (completely healthy) to 1 (full-blown mastitis). The main model input was lactate dehydrogenase (LDH; mumol/min per L) x milk yield. Test data containing 253 mastitis cases were used. Proportional samples were collected from each cow at each milking and analyzed for LDH and somatic cell count (SCC). The basis for the health definitions was veterinary treatment records. A refinement of the basic health definitions was made using systematic positive deviations in log(SCC) to indicate untreated infections. Two subsets of cows were identified: mastitic cows and cows completely free of mastitis (healthy controls). The time-profiles of these 2 groups in a 60-d window relative to day of veterinary treatment were examined. Model reliability throughout all stages of lactation and degrees of infection was examined using SCC as a continuous measure of degree of mastitis. The time-profile for the health controls was flat throughout the 60-d window with a median MR of 0.02. In contrast, the profile of the mastitic cows increased above the control cows' baseline from about -6 d, rising to a MR value of 0.20 at d 0, and declining to the control level after treatment. There were significant differences between mastitic and healthy cows from -4 to +2 d relative to veterinary treatment. When cases were time-aligned to peak of infection, rather than veterinary treatment, there was a much sharper peak to the time-profile of mastitic cows. The median MR at peak was 0.62 and the mean was 0.80. Using these data, the MR value of 0.62 had a <1% likelihood of actually coming from a healthy control. Testing against SCC, on the whole data set, showed that only 2.1% of all MR values had an error >0.7. These estimates of model reliability are comparable with the greatest values reported in the literature and, additionally, the model was able to detect significant differences between mastitic and healthy cows 4 d before treatment. It was also found that specificity/sensitivity calculations are inappropriate for evaluating time-related changes and a progressive scale of predicted mastitis state.