Data-based monitoring of dairy cows - milk progesterone as a mirror of fertility

With an annual production value of 639 million euros, dairy production is the fourth most important agricultural sector of Flanders. About 18% of all farms have dairy cows, summing up to a total of more than 300 000 animals. The modern dairy sector is characterized by strong specialization and scale enlargement, in which technology plays a prominent role. This technology supports the management on farm, and secures that despite the large herd sizes, individual animals can remain monitored. Although several technological developments are already commercialized and implemented on farm, still many challenges remain. Not in the least, there is a challenge of interpretation: likewise all other biological systems, dairy cows show a large individuality and variability, which renders it complex and challenging. Certainly in the case of monitoring systems for individual animals to support decision making on farm, there is a need for smart, physiology-based interpretation of the sensor data. Bad reproduction performance is the second most important cause of economic losses on Flemish dairy farms, summing up to on average 49€ per cow per year. Crucial to this aspect of dairy farming is the correct and timely identification of the cows' fertility status. Correct detection of estrus, onset of cyclicity and pregnancy allow to optimize (re-)insemination and treatments, which in their turn contribute to shorter calving intervals. Classically, estrus detection is done via visual observations of external symptoms or by using technology to continuously monitor increased restlessness in the estrous period. Although these technologies have shown their merit, they are not capable of identifying onset of cyclicity, ovarian problems or pregnancy. In the last decades, recent developments have led to technology for the automated measurement of milk progesterone on farm. Milk progesterone, in contrast to behavior-based technology, reflects the presence or absence of a corpus luteum on the ovaries. Monitoring the dynamics of the progesterone concentration therefore has the potential to provide a more complete image of a cow's fertility status. To this end, clear and consistent interpretation of the progesterone time series, while working in a cost-efficient and automated setting is essential. More concretely, the progesterone data should be converted to solid actions, while taking the variability caused by individuality of the animals, but also by measurement errors into account. The main objective of this PhD work was to develop an online monitoring algorithm based on milk progesterone, capable of unambiguously discriminating between the different fertility statuses, which could be implemented on farm and work in a cost-effective setting. More specifically, this could be translated into following sub goals: (1) mathematical characterization of the milk progesterone concentration, taking into account the physiological background of luteal dynamics; (2) integration of the mathematical model into an on-line monitoring system implementable on farm; (3) investigation of the link between the luteolytic drop in milk progesterone and ovulation time in order to improve insight in the optimal insemination window; and (4) benchmark the proposed methodology against the current state of the art for progesterone monitoring. By investigating both the sensitivity and specificity of the algorithms, but also looking into robustness for missing samples during the crucial moment of luteolysis, this work was further validated. Growth and regression of the corpus luteum are associated with a steady increase and a sudden decrease of progesterone. Each of these dynamics can be linked to a specific reproduction status and described with a mathematical function. During the postpartum anestrus phase after calving, no progesterone is produced and a constant suffices to characterize baseline height. Once cyclicity commenced, the successive increases and decreases can accurately be described with a sigmoidal Hill and Gompertz function respectively. After insemination and when successful conception resulted in the establishment of pregnancy, the development of a pregnancy corpus luteum and its associated increase in progesterone again can be characterized with the Hill function. These functions have the advantage that the lengths of the follicular and luteal phases might vary, even as the slopes of the increases and decreases, while still maintaining the general physiology-based shape of each cycle. Another important advantage of these functions is that they can easily be implemented in an on-line monitoring system. More specifically, the general principle is to detect onset of cyclicity via the deviation of the progesterone concentration from the baseline. Next, the increasing Hill function is fitted, and updated each time a new measurement becomes available. The residual of this new measurement with the fitted function is evaluated via a statistical process control chart in order to detect large negative deviations from the fitted luteal concentration. If there is enough evidence for a decrease in progesterone, the decreasing Gompertz function is added. At this time, the mathematical model can be used to calculate several model-based indicators, which allows to characterize the cycle and moment of luteolysis independently of the real sampling rate. Once luteolysis is detected, it is important to know when the succeeding ovulation will follow in order to optimize timing of insemination and maximize the chance on conception. The high workload and costs to investigate this, make this research rather difficult. Two studies were conducted to determine timing of ovulation after luteolysis. In the first, the ovarian status of the cows was synchronized and ultrasonography used to detect ovulation during the second estrus after synchronization. For the second study, the cows were not synchronized but the preovulatory LH surge was taken as a proxy for ovulation. In the latter study, also the alerts raised by the visual observation of estrus symptoms and an activity-sensor system were included. Progesterone-based systems were more sensitive and had a higher positive predictive value than visual observation of estrus symptoms and the activity based sensors. Moreover, it was shown that ovulation and the LH surge follow on average 77 ± 10 hours and 62 ± 12 hours after luteolysis detected by the developed algorithm. Using the indicators derived from the mathematical model describing the estrous cycle, a more consistent relation with the LH surge was found. However, as until now the sampling frequency was very high, the true value of these indicators was not yet exposed. To further validate the developed system and to compare this with the current state of the art for monitoring milk progesterone, a large dataset with all possible (combinations of) progesterone patterns was required. Therefore, a reproduction function model coupled to a lifetime performance model was employed to simulate scaled progesterone profiles, and a technique to convert these scaled profiles into realistic curves as they were measured on farm, was developed. Different sampling schemes, from which the reduced sampling scheme mimicked the case of missing values during luteolysis, were applied to this dataset. The sensitivity, specificity and robustness of the alerts was investigated and compared to a multiprocess Kalman filter in combination with a fixed threshold, which is the currently implemented tool to interpret the progesterone values on farm. We showed that the developed algorithm captured luteolysis almost simultaneously with the simulated luteolysis, in contrast to the Kalman filter for which the alerts came on average 1 to 2 milkings later. Moreover, the analysis provided evidence that alerts based on a model-derived indicator taking the height and baseline of the curves into account, allowed to estimate moment of luteolysis consistently, independent of missing samples. Monitoring milk progesterone allows to identify the different fertility status of lactating dairy cows on farm. However, to optimize farm management, not only identification is important, but also the concrete actions provided to the farmers. Unfortunately, whether or not it is the 'best' choice to inseminate, to treat or to cull is dependent on more than only the cows: ideally, also the physiological background of the animals, the economic environment, the farm system and so on should be taking into account. This type of overall data-integration for the optimization of decision support was identified as the main topic for future research.

Jaar van publicatie:2018

Toegankelijkheid:Open