Machine learning modeling to predict causes of infectious abortions and perinatal mortalities in cattle

A plethora of infectious and non-infectious causes of bovine abortions and perinatal mortalities (APM) have been reported in literature. However, due to financial limitations or a potential zoonotic impact, many laboratories only offer a standard analytical panel, limited to a preestablished number of pathogens. To improve the cost-efficiency of laboratory diagnostics, it could be beneficial to design a targeted analytical approach for APM cases, based on maternal and environmental characteristics associated with the prevalence of specific abortifacient pathogens. The objective of this retrospective observational study was to implement a machine learning pipeline (MLP) to predict maternal and environmental factors associated with infectious APM. Our MLP based on a greedy ensemble approach incorporated a standard tuning grid of four models, applied on a dataset of 1590 APM cases with a positive diagnosis that was achieved by analyzing an extensive set of abortifacient pathogens. Production type (dairy/beef), gestation length, and season were successfully predicted by the greedy ensemble, with a modest prediction capacity which ranged between 63 and 73 %. Besides the predictive accuracy of individual variables, our MLP hierarchically identified predictor importance causes of associated environmental/maternal characteristics of APM. For instance, in APM cases that happened in beef cows, season at APM (spring/summer) was the most important predictor with a relative importance of 24 %. Furthermore, at the last trimester of gestation Trueperella pyogenes and Neospora caninum were the most important predictors of APM with a relative importance of 22 and 17 %, respectively. Interestingly, herd size came out as the most relevant predictor for APM in multiparous dams, with a relative importance of 12 %. Based on these and other mix of predicted environmental/maternal and pathogenic potential causes, it could be concluded that implementing our MLP may be beneficial to design a more cost-effective, case-specific diagnostic approach for bovine APM cases at the diagnostic laboratory level.