Prediction of the reproductive status of cattle on the basis of milk progesterone measures: model description

Reproductive management, in particular timely oestrus detection, is important for profitable dairy production. The aim of this study was to develop a biological model to predict reproductive state on the basis of milk progesterone measures. A number of additional inputs were incorporated to make use of other known effectors of reproductive performance that are not reflected in progesterone levels. These are: days from calving, breed, parity, signs of behavioural oestrus, insemination dates, pregnancy determinations, energy status, body fat status, milk urea content and reproductive disorders associated with calving. A dynamic, deterministic model was developed. It is designed to run each time a new trigger input (progesterone, behavioural oestrus, inseminations, pregnancy determinations) occurs using the current and previous values and can run in the absence of the additional inputs. The milk progesterone values are smoothed using an extended Kalman filter before being processed in the biological component of the model. The model predicts the reproductive status of the cow, which can be one of three mutually exclusive states: postpartum anoestrus, oestrus cycling, and potentially pregnant. The other model outputs are all reproductive status specific with the exception of days to next sample (DNS), which is calculated in each model run regardless of reproductive status. DNS is designed to feedback to the sampling system so that the frequency of milk sampling (i.e. progesterone measurement) can be varied according to the predicted likelihood of a future reproductive event, such as onset of oestrus cycling. The other model outputs are: risk of prolonged postpartum anoestrus, risk and type of ovarian cyst, onset of oestrus, likelihood of a potential insemination succeeding, and likelihood of being pregnant (following oestrus). The model was evaluated using three simulated datasets consisting of a timeseries of progesterone values centred on each of the three reproductive statuses and including relevant additional information. Test runs were carried out on the full datasets and then on reduced data. The data reductions were made by using only those values that would have been available if the model days to next sample function was used to control sampling frequency. The sensitivity of the model to noise in the raw progesterone data was examined by adding 1, 2, or 3 residual standard deviations (1.85 ng/ml) random variation to the original data and evaluating model performance. The model was found to be able to readily identify and distinguish reproductive states. A reduction in sampling frequency to 36% of original sample resulted in an average increase in days to detection of oestrus of 0.36. The addition of 1 S.D. noise did not cause additional oestruses to be detected and all oestruses were correctly identified. However, when 2 or 3 S.D. noise were added, the model found on average 1.4 and 3 extra oestruses. It was concluded that reproductive status can be predicted from milk progesterone values using a biological model and that such a model is robust to reductions in sampling frequency number and to a doubling in the random variation in the raw progesterone values. It therefore has the potential to provide the basis for a useful reproductive management tool.