The conventional window for ultrasonic pregnancy diagnosis in sows is 22–25 days post-insemination, which often results in missed opportunities for the optimal re-insemination of non-pregnant sows and elevated production costs. This present study aimed to establish an early pregnancy detection method for sows at 12–18 days post-insemination, thereby providing a reference for efficient reproductive management. Saliva, urine and vaginal secretions were collected from sows during this period, and seven metabolites were quantified. Seven machine learning models were employed for data analysis, after which the optimal combination was determined, and the detection protocol was refined using recursive feature elimination. The results revealed that the majority of metabolites in saliva and urine differed significantly between pregnant and non-pregnant groups (p < 0.05). Among the models evaluated, the random forest algorithm exhibited the best predictive performance, with accuracy ranging from 0.59 to 1.00. Saliva sampled at 17 days post-insemination was identified as the optimal diagnostic medium, and 100% prediction accuracy was achieved by measuring only three metabolites: Glc, Ste, and Xan. The diagnostic approach established in this study allows pregnancy detection 5–8 days earlier than conventional methods, with the additional benefits of non-invasive sampling and minimal stress to sows. Accordingly, it provides a novel reference for enhancing the efficiency of swine production.
Feng et al. (2026) studied this question.