The authors report on the use of a sampling devitz to collect the breath from individual members of a herd of dairy cattle during a two-week period. The response of an array of six semiconducting oxide gas sensors to the breath samples has been recorded and subsequently modelled by a time-dependent, linear, second-order system. Four characteristic sensor parameters have been estimated using a neural network, and these parameters have been used to train a predictive multilayer perceptron network. The results show that either a static response parameter (based on the difference in the signal from zero time) or a single time constant can be used to predict reasonably well the health of the cow as judged against blood samples. In both cases, the identification rate of unknown samples being about 76%. Further improvements may be possible through the use of network compensation of variations in sample temperature and humidity.