Predicting mastitis in dairy cows using neural networks and generalized additive models: A comparison

The aim of this paper is to develop and compare methods for early detection of oncoming mastitis with automated recorded data. The data were collected at the Danish Cattle Research Center (Tjele, Denmark). As indicators of mastitis, electrical conductivity (EC), somatic cell scores (SCS), lactate dehydrogenase (LDH), and milk yield are considered. Each indicator is decomposed into a long-term, smoothed component, and a short-term, residual component, in order to distinguish long-term trends from short-term departures from these trends. We also study whether it is useful to derive a latent variable that combines residual components into a score to improve the model. To develop and verify the model, the data are randomly divided into training and validation data sets. To predict the occurrence of mastitis, neural network models (NNs) and generalized additive models (GAMs) are developed using the training set. Their performance is evaluated on the validation data set in terms of sensitivity and specificity. Overall, the performance of NNs and GAMs is similar, with neither method appearing to be decisively superior. NNs appear to be marginally better for high specificities. NNs model results in better classification with all indicators, using individual residuals rather than factor scores. When SCS is excluded, GAMs shows better classification result when milk yield is also excluded. In conclusion, the study shows that NNs and GAMs are similar in their ability to detect mastitis, a sensitivity of almost 75% observed for 80% of fixed specificity. Including SCS in the models improves their predictive ⩾5% ability.