Mastitis detection in dairy cows by application of neural networks

The aim of the present research was to investigate the usefulness of neural networks (NN) in the early detection and control of mastitis in cows milked in an automatic milking system. A data set of 403,537 milkings involving 478 cows was used. Mastitis was determined according to two different definitions: udder treatment or somatic cell counts (SCC) over 100,000/ml (1) and udder treatment or SCC over 400,000/ml (2). Mastitis alerts were generated by an NN model using electrical conductivity, milk production rate, milk flow rate and days in milk as input data. To develop and verify the model, the data set was randomly divided into training and test data subsets. The evaluation of the model was carried out according to block-sensitivity, specificity and error rate. When the block-sensitivity was set to be at least 80%, the specificities were 51.1% and 74.9% and the error rates were 51.3% and 80.5% for mastitis definitions 1 and 2, respectively. Additionally, the average number of true positive cows per day ranged from 1.2 to 6.4, and the average number of false negative positive cows per day ranged from 5.2 to 6.8 in an average herd size of 24 cows per day for the test data. The results for the test data verified those for the training data, indicating that the model could be generalized. The performance of the NN was not satisfactory. A decrease in the error rate might be achieved by means of more informative parameters.