Data-driven insights into priority predictors associated with conception success in indigenous goats in Uganda

Abstract Identifying predictors associated with conception success in indigenous goats is important for prioritising reproductive management in smallholder livestock systems. This observational study applied five machine learning algorithms, Logistic Regression (LR), Random Forest (RF), XGBoost (XGB), Support Vector Machine (SVM) and Artificial Neural Network (ANN), to 900 doe-mating records from 240 smallholder farms across four agro-ecological zones in Uganda, of which 68% resulted in confirmed pregnancy. Twenty-five predictors spanning animal, health, management and environmental domains were evaluated. Feature importance was quantified using standardised log-odds coefficients for LR, normalised Gini impurity reduction for RF and XGB, and permutation importance for SVM and ANN, and cross-algorithm agreement was assessed using Kendall’s W. On the held-out test set ( n = 180), LR achieved the highest discrimination (ROC-AUC 0.691, 95% CI 0.617–0.764) and the best calibration (Brier score 0.221). RF and XGB showed higher sensitivity (0.910 and 0.869) but low specificity (0.241 and 0.293), while SVM and ANN performed close to chance (AUC 0.540 and 0.519). Overall discrimination was modest, so the findings should be interpreted as associations rather than as a validated predictive tool. Body condition score (BCS) at mating was the most consistently ranked predictor, ranking first in four of the five algorithms, with an LR odds ratio of 1.40 per standard deviation increase, corresponding to approximately 40% higher odds of conception. Rainfall (CHIRPS) and NDVI (MODIS) ranked among the top predictors in the tree-based ensemble models, while parity, disease events within 30 days of mating, Temperature–Humidity Index (THI) and FAMACHA endoparasite score were negatively associated with conception in the LR model. Kendall’s W of 0.48 ( p < 0.01) indicated moderate cross-algorithm agreement across the full predictor ranking. Domain-level analysis for LR, RF and XGB attributed 51–59% of total importance to animal-level predictors, 22–31% to environmental predictors, and 11–22% to health and management predictors. These associations are observational and do not establish causality. Within this dataset, pre-mating body condition, strategic deworming, timing of mating with pasture availability, and mitigation of heat stress emerge as priority areas for investigation, pending experimental validation. Predictors not captured in this study, including diet composition, forage type, mineral supplementation, grazing behaviour and water access, are acknowledged as limitations.