Auswertung serieller Daten zur Mastitiserkennung mit Hilfe der lokalen Regression

The aim of this study was to detect incidence of mastitis in an automatic milking system using serial information. Data from 112,000 milkings of the research dairy herd Karkendamm of the University of Kiel were available. The incidence of mastitis was defined both on therapies carried out and on weekly somatic cell count measurements. The time series of electric conductivity of quarter milk were analysed to find deviations as a sign for mastitis. Three methods were performed to find mastitis cases. First, a local regression method using the SAS-procedure LOESS, second a moving average system and third an exponentially weighted moving average were applied. The used methods provided similar results. The goodness of alerts varied dependent on the threshold value. A low threshold (3%) led to a sensitivity of nearly 100%, however the specifity was only about 30% and thus the error rate was high (about 70%). With increasing thesholds (7%) sensitivity decreased to 70% and specifity increased to 75%. Error rate was slightly reduced to 60%.