Validation of a novel milk progesterone-based tool to monitor luteolysis in dairy cows. Timing of the alerts and robustness against missing values

ABSTRACT

Automated monitoring of fertility in dairy cows using milk progesterone is based on the accurate and timely identification of luteolysis. In this way, well-adapted insemination advice can be provided to the farmer to further optimize the fertility management. To properly evaluate and compare the performance of new and existing data-processing algorithms, a test dataset of progesterone time-series that fully covers the desired variability in progesterone profiles is needed. Further, the data should be measured with a high frequency to allow rapid onset events, such as luteolysis, to be precisely determined. Collecting this type of data would require a lot of time, effort and budget. In the absence of such data, an alternative was developed using simulated progesterone profiles for multiple cows and lactations, in which the different fertility statuses were represented. To these, relevant variability in terms of cycle characteristics and measurement error was added, resulting in a large cost-efficient dataset of well-controlled but highly variable and farm-representative profiles. Besides the progesterone profiles, information on (the timing of) luteolysis was extracted from the modelling approach and used as a reference for the evaluation and comparison of the algorithms. In this study, two progesterone monitoring tools were compared: a multiprocess Kalman filter combined with a fixed threshold on the smoothed progesterone values to detect luteolysis, and a progesterone monitoring algorithm using synergistic control ‘PMASC’, which uses a mathematical model based on the luteal dynamics and a statistical control chart to detect luteolysis. The timing of the alerts and the robustness against missing values of both algorithms were investigated using two different sampling schemes: one sample per cow every eight hours versus one sample per day. The alerts for luteolysis of the PMASC algorithm were on average 20 hours earlier compared to the ones of the multiprocess Kalman filter, and their timing was less sensitive to missing values. This was shown by the fact that, when one sample per day was used, the Kalman filter gave its alerts on average 24 hours later, and the variability in timing of the alerts compared to simulated luteolysis increased with 22%. Accordingly, we postulate that implementation of the PMASC system could improve the consistency of luteolysis detection on farm and lower the analysis costs compared to the current state of the art.