Alternative modelling techniques in oestrus detection for dairy cows

Cow activity is the best linked parameter to the cow oestrus state. Numerous studies have been carried out to test the reliability of the system aimed at detecting oestrus, producing generally positive results, though not always unanimous. In fact if performances are measured in terms of efficiency, taken as the relationship between detected oestrus and total oestrus, results are obtained which are analogous, or even more reliable than those obtained through careful visual observation. However, an increase in efficiency corresponds to a reduction in specificity, that is, the ratio of the number of true negative values to the total of true negative and false positive values. Besides a problem of poor accuracy in signalling the oestrus, there is also a problem of possible delay in observing the alarm signal with respect to the effective occurrence of the phenomenon, due to the fact that the activometers are usually only read during milking and therefore about every twelve hours. Since the probability of successful insemination is at its peak at around 12 hours from the onset of the oestrus (Maatje and Loeffler, 1998), and considering that insemination can also take place several hours after discovery of an alarm, it is desirable for the signalling of the increase in activity to be as prompt as possible and therefore with readings at reduced time intervals. For this reason, a system which allows for the reading of activity values at hourly intervals has recently been introduced on the market, and it is on this system that part of our research has been focused. The importance of implementing the oestrus detection system based on activometer readings is therefore recognised, both with regards to the methods for obtaining the data, and, above all, with regards to the techniques for elaborating the data. One first implementation possibility consists, as we have said, of an increase in the frequency of reading activity values. That this solution leads to an improvement in the accuracy of signalling has been demonstrated, besides by the above-cited work, also by the studies of Pulvermacher and Maatje (1992) and by Zappavigna et al. (2000). A second possibility lies in the combination of the activity parameter with other indicators, in a multivariate way. A useful parameter for our purposes seems to be milk production and flux, for which there is a reduction in correspondence with the oestrus (Schofield, Phillips and Owens, 1991; Schlünsen, Schön and Roth, 1988). Although the accuracy of this indicator appears to be much poorer than activity parameters, this data is interesting for its use in combination with other data, given the fact that it is easily acquired. The possibility of improving oestrus detection through a combination of several indicators using quite complex algorithms has been investigated in particular by de Mol. et al (1993-2000), by Scholten et al. (1995), and by Krieter et al. (2002). As a result, it will be one of the subjects of this work.