Evaluation of machine learning predictions for early reproductive success in commercial US dairies

ABSTRACT: Reproductive performance affects the profitability of a dairy herd. The ability to understand the reproductive capabilities of individual cows and the use of targeted reproductive management could optimize reproductive performance of dairy herds. To address this need, the early reproductive success prediction was developed using a light gradient-boosting machine algorithm, which included herd-level reproduction data, weather data, genomic-enhanced predicted transmitting ability, individual cow information, milk production, health events, and previous lactation performance data. The objective of this retrospective study was to evaluate the ability of the early reproductive success algorithm to predict pregnancy probability by 110 DIM at 43 DIM, which was before the end of the voluntary waiting period for enrolled herds. The study included 9,969 Holstein and 9,464 Jersey multiparous cows that calved in 2022 from 7 US commercial herds. Cows were ranked by their predicted probability within their own herds and then assigned to deciles based upon this ranking. Cows' reproductive and herd exit data were collected for 18 mo following calving from on-farm management software. Data were analyzed using a generalized linear mixed model. The worst 10% and best 10% early reproductive success prediction deciles were different for pregnancy at first insemination (25.3% vs. 44.2%), proportion pregnant at 110 DIM (35.6% vs. 64.8%), and proportion of cows that gave birth to a live calf to initiate the following lactation (49.1% vs. 77.8%), with percentage improvements in performance of 75%, 82%, and 58%, respectively. The predicted worst 10% and best 10% deciles were different for abortion incidence (20.9% vs. 6.8%) and whether cows were sold within enrollment lactation (43.8% vs. 17.2%), with percentage improvements in performance of 67% and 61% for these 2 metrics. These results demonstrated the ability of the early reproductive success algorithm to predict differences in pregnancy per insemination for all services, abortion incidence, proportion of cows sold in enrollment lactation, and proportion of cows producing a live calf. Further research is needed to determine whether the early reproductive success prediction has potential to be used to help dairy producers develop targeted reproductive management strategies.