Analysing serial data for mastitis detection by means of local regression

The aim of this study was to assess the potential of detecting mastitis in an automatic milking system using serial information from electrical conductivity of milk. Data from 160 cows from the experimental dairy farm “Karkendamm” of the University of Kiel were available over a period of 14 months. The reference data for incidence of mastitis were treatments (or visual observation of clinical mastitis signs) and the weekly milk somatic cell count (SCC) measurements of all cows. Samples of SCC exceeding 400,000 and 100,000 cells/ml were used as two boundaries together with treatments to define cases of mastitis.The time series of electrical conductivity of quarter milk were analysed to find significant deviations as a sign for mastitis. Three statistical methods were tested: (1) a moving average, (2) an exponentially weighted moving average and (3) a locally weighted regression. Alerts for mastitis were given when the relative deviation between the measured value and the estimated value exceeded a given threshold value, expressed as a percentage. The three methods provided similar results regarding sensitivity, specificity and error rate. The reliability of alerts varied depending on the threshold value. A low threshold (3%) led to a sensitivity of nearly 100%, however, the specificity was only about 36% and thus the error rate was high (about 70%). Increasing the threshold up to 7% decreased sensitivity to 70% and increased specificity to 84%. In this case, the error rate was slightly reduced to 60%. The three methods showed a good sensitivity and specificity for an appropriate threshold value, but also a high error rate. In the present study, the moving average was the simplest method to detect mastitis and the other methods showed no advantage related to it.