The use of machine learning methods to predict sperm quality in Holstein bulls

Production of high quality bull semen is important for a successful breeding program. Therefore, predicting sperm quality of production bulls would be a useful tool to help identify and manage environmental factors that influence sperm quality. The successful implementation of a prediction tool could help artificial insemination (AI) centers to adjust management factors, depending on their influence on sperm quality, to ensure stable sperm production and to improve quality of the produced semen throughout the year.

In the past, genetic and environmental factors have been found to have an influence on the quality and quantity of bull semen [[1], [2], [3]]. A longer collection interval (the number of days between two semen collections) has a positive influences on ejaculate volume, sperm concentration and sperm output [1,2,4]. Additionally, the collection technicians (bull handler and semen collector) had a significant effect on ejaculate volume and sperm output [1]. Another important factor reported to affect sperm quality and semen production is the age of bull [1,5] in which case sperm quality and sperm output increase with the bull's age. Different mounting practices can also influence sperm production. For example, prior to semen collection, the bull can be false mounted or actively restrained. Both methods increase the total number of sperm per ejaculate [[6], [7], [8]].

An increasing emphasis has been put on the effect of climate on sperm quality. Heat stress, either directly on the scrotum by means of scrotal insulation [9] or by climatic conditions influencing the whole animal [10,11] has been shown to have a negative impact on sperm quality and quantity. Relatedly, the effect of seasonal changes has been inconclusive, with some studies considering the effect of season to be significant [12,13], whereas other studies have failed to prove any seasonal effects [14]. Finally, the temperature on the day of semen collection as well as the temperature during the time of spermatogenesis has been found to influence sperm quality [15].

Given that a vast range of factors contribute to sperm quality, with their effect likely being nonlinear or interactive with other factors, modeling sperm quality is assumedly prone to omitted variable errors. Machine Learning (ML) algorithms do not require a complete understanding of the underlying complex mechanism in order to build a model, as they learn the parameters that optimize prediction accuracy from the data. They, therefore, offer an interesting alternative for deriving prediction models and have been effectively applied to predict the conception success of dairy cows [16], as well as the fertilizing capacity of ejaculates from bucks [17], boars [18] and humans [19].

This study uses the ML approach for the prediction of bull sperm quality. For this, we used the Lasso [20] and Gradient Boosting [21] methods, which manage to identify from a given set of potential influencing variables those with the highest joint prediction power. While such ML approaches offer great flexibility in choosing variables for predictions models, they often lack interpretability regarding how the relationships between factors work. To better understand the relationship of explanatory variables, another ML method called Group Lasso [22] was evaluated. This method allows the examination of the relationship between groups of variables affecting sperm quality with statistical inference. The resulting network diagram can serve as an input for further investigation with theory-based mechanistic approaches. As it is shown in the results section of the paper, this method, albeit more restrictive in variable selection than Lasso and Gradient Boosting, often performs similar or even better with respect to prediction accuracy. This might also indicate that flexible ML methods that select variables without reasoning about their relation to each other might lead to models that overfit the training data and are therefore less generalizable.

留言 (0)

沒有登入
gif