Integrating machine learning and SHAP interpretations as a decision-support tool for embryo quality assessment in...

Embryo quality is a fundamental determinant of reproductive success and genetic progress in cattle, yet traditional morphological grading remains inherently subjective and limited in predicting actual pregnancy outcomes. This study proposes an explainable artificial intelligence (XAI) framework to predict in vivo-derived embryo quality in dairy cattle while providing transparent, biologically aligned insights into model decision-making. Utilizing a comprehensive dataset of 418 embryo records from 389 lactating Holstein cows, 27 attributes-including various hormonal levels, follicle sizes, and body condition scores-were integrated into the machine learning analysis. Six models, specifically Random Forest, K-Nearest Neighbors, XGBoost, LightGBM, AdaBoost, and Histogram-Based Gradient Boosting (HistGradientBoost), were implemented and evaluated using accuracy, recall, precision, and F1-score. Interpretability was established through SHAP (SHapley Additive exPlanations) to provide both global patterns of feature importance and local, case-specific explanations. Among the models, HistGradientBoost demonstrated superior performance, achieving an accuracy of 71.1% and an F1-score of 0.715. Global SHAP analysis revealed that hormonal profiles-specifically progesterone dynamics (p4govs, p4pgovs, p4ai, p4flush) and estradiol at insemination (e2ai)-were the dominant predictors of embryo quality. High-quality embryos were associated with optimal hormonal trajectories and elevated accessory spermatozoa counts (asc). Furthermore, local SHAP waterfall plots demonstrated how individual features, such as post-GnRH progesterone, drive individual quality predictions for specific samples. By bridging computational modeling and reproductive biology, this framework provides actionable decision-support tools for veterinarians to refine synchronization protocols and optimize reproductive outcomes in modern cattle breeding systems.