Reproductive performance is the beating heart of dairy profitability. A cow that conceives quickly after calving maintains a short calving interval, spends more of her life producing milk at peak levels, and costs her owner less in feed, labor, and veterinary care. A cow that lingers open, by contrast, drains resources month after month. Researchers have long known that a constellation of factors influences this interval, including the age of the animal and a suite of postpartum diseases such as metritis, mastitis, ketosis, retained placenta, and ovarian cysts. Epidemiological studies stretching back decades have documented, for example, that metritis reduces conception rates and prolongs the interval to subsequent pregnancy, while subclinical ketosis delays the return of estrus and extends the time from calving to both first insemination and conception.

The problem with conventional machine learning in this domain has been trust. Powerful algorithms such as gradient boosting ensembles and deep neural networks can capture nonlinear relationships among predictors that traditional regression models miss, but they do so inside a black box. A veterinarian handed a prediction has no way of knowing whether the model is responding to a genuine biological signal or to some artifact of the data. This tension between accuracy and interpretability has fueled the rise of explainable artificial intelligence, or XAI, a family of techniques designed to pry open the decision-making machinery of complex models. Among these techniques, the most theoretically robust is SHAP, short for SHapley Additive exPlanations, which borrows its mathematics from cooperative game theory.

SHAP treats every input variable as a player in a game where the payout is the model’s prediction. By systematically evaluating how the prediction changes as each feature is added or removed across all possible coalitions of features, SHAP assigns each variable a value quantifying its marginal contribution. For an individual cow, this means the method can say precisely how much her age, her history of mastitis, or her bout with ketosis pushed the predicted interval to first insemination up or down relative to the average. Aggregated across the whole dataset, these local explanations yield a global ranking of feature importance. Crucially, the method satisfies axioms of local accuracy, consistency, and missingness that competing explanation techniques such as partial dependence plots and LIME do not simultaneously guarantee, which is why it has become the interpretability tool of choice in animal science.

In the new study, a team led by Elif Çelik Gürbulak of Erciyes University assembled 466 observational records from a dairy cattle farm, later refined to 439 complete cases after excluding heifers and records with missing values. The outcome variable was the number of days from calving to first insemination, with shorter intervals signaling better reproductive performance. The predictors were age and binary indicators for five common postpartum conditions. The researchers fitted an eXtreme Gradient Boosting, or XGBoost, regression model, tuning its hyperparameters through a grid search combined with five-fold cross-validation across 432 parameter combinations. The optimal configuration used a maximum tree depth of two, a learning rate of 0.05, and subsampling rates of 0.7, with 100 boosting rounds. Interpretability was then layered on top using the TreeSHAP algorithm, which computes exact Shapley values for tree-based models in polynomial rather than exponential time.

The global picture that emerged from the SHAP analysis was strikingly coherent with veterinary intuition. Age dominated the model, posting the highest mean absolute SHAP value of 3.205 and accounting for nearly ninety percent of the model’s gain in the XGBoost importance metrics. Older cows contributed strongly to predictions of longer intervals to first insemination, a pattern consistent with earlier findings that conception probability declines with increasing parity and that pregnancy rates fall after the fourth lactation. Mastitis ranked second with a mean absolute SHAP value of 0.324, followed by retained placenta at 0.173, ovarian cysts at 0.079, ketosis at 0.073, and metritis at 0.072. The positive SHAP contributions of the postpartum diseases indicated that their presence pushed predictions toward longer intervals, aligning with known biology in which mastitis pathogens trigger prostaglandin release, disrupt the estrous cycle, and impair oocyte quality and early embryonic development.

When the researchers stratified the herd into low, medium, and high reproductive performance classes using percentile thresholds, the disease burden told a clear story. Mastitis prevalence climbed progressively from 19.6 percent in the high-performance group to 38.4 percent in the low-performance group, while retained placenta affected 21.2 percent of the lowest performers. The mean interval to first insemination stretched from 56.27 days in the top group to 78.47 days in the bottom group. Radar charts visualizing the ten best and ten worst animals reinforced the gradient, showing that higher overall disease prevalence accompanied poorer reproductive outcomes. At the level of individual cows, the SHAP values varied in both direction and magnitude, revealing heterogeneity that a single herd-level statistic would have concealed entirely.

Yet the study is equally notable for what it honestly admits it cannot do. On the independent test dataset, the model’s predictive performance collapsed: the root mean square error rose to 39.05 days, the mean absolute error to 18.58 days, and the coefficient of determination fell to a mere 0.01, meaning the model explained essentially none of the variance in unseen animals. The training set metrics, with an R-squared of 0.18, were only modestly better. The authors are explicit that the SHAP importance values should be read as explanations of the patterns the fitted model learned, not as definitive evidence of causal or universally generalizable biological relationships. Important variables influencing reproduction, such as nutrition, genetics, environment, and management practices, were either absent from the dataset or insufficiently represented, and the modest sample size of a single farm further limits external validity.

This candor is precisely what makes the study a useful template rather than an overhyped failure. The authors frame the work as a methodological proof of concept, illustrating how XGBoost and SHAP can be integrated into veterinary data analysis before any clinical decision-support application is attempted. Their conclusion echoes a growing consensus in the explainable AI literature: a model should be judged not only by its predictive accuracy but by the quality and trustworthiness of its explanations, and transparent attribution of feature contributions is a prerequisite for any tool a veterinarian might actually rely on at the bedside, or in this case at the stall side. Future iterations, the researchers argue, will need larger and more diverse herds, a broader range of reproductive, nutritional, genetic, and environmental variables, and validation across multiple farms before such models can inform real herd-management decisions.

The broader significance extends well beyond dairy science. As machine learning spreads through livestock production, companion animal medicine, and epidemiology, the demand for models that can justify their predictions will only intensify. Studies applying SHAP have already illuminated infection risk in cats, mortality prediction from veterinary health records, and survival in equine colic, and comparisons with traditional logistic regression continue to clarify where each approach belongs: regression for causal interpretation, SHAP-augmented machine learning for clinical interpretation of complex patterns. The Turkish team’s dairy cow study, with its transparent accounting of both strengths and limitations, offers a candid glimpse of that future, one in which the algorithms advising farmers and veterinarians can finally show their work.

Subject of Research: Explainable machine learning assessment of reproductive performance and disease risk factors in dairy cows

Article Title: Assessment of reproductive performance in dairy cows using explainable machine learning

Article References: Çelik Gürbulak, E., Kara, U., Canooğlu, E., Yüceel, H. A., Çetin, E., Demirel, M., & Gürbulak, K. (2026). Assessment of reproductive performance in dairy cows using explainable machine learning. Archives Animal Breeding, 69(3), 469-476. https://doi.org/10.5194/aab-69-469-2026

Image Credits: AI Generated

DOI: 10.5194/aab-69-469-2026

Keywords: dairy cows, reproductive performance, explainable artificial intelligence, XGBoost, SHAP, machine learning, veterinary science, mastitis, ketosis, retained placenta, metritis, ovarian cysts