Comparison of analysis techniques for on-line detection of clinical mastitis.

Three techniques were compared for analysis of automatically collected data from the milking parlor. Mammary quarters showing signs of clinical mastitis were compared with randomly selected healthy quarters. Automatic data were analyzed from the milking on which the milkers observed clinical mastitis as well as data from the two prior milkings. Electrical conductivity of milk was not corrected for individual cows. Milking parlor data were preprocessed so that information on the electrical conductivity pattern during a milking was retained. Principal component analysis was used to verify whether variation in the data was caused by mastitis. Performance of logistic regression models for detection of clinical mastitis was compared with that of backpropagation neural networks. Variation in the quarter data was caused by mastitis. Automatic data from infected quarters did not always differ from data from healthy quarters, especially from the two prior milkings. The detection performance of the logistic regression model was similar to that of the neural networks. When both models were tested on the development data, sensitivity was approximately 75%, and specificity was approximately 90% at the milking of mastitis observation. Detection results were lower for the prior milkings. Therefore, not all incidences of clinical mastitis cases could be detected before clinical signs occurred.