National-scale evaluation of groundwater fluoride in Ethiopia using spatially buffered machine learning

Excess groundwater fluoride (F⁻) is a public health concern in the East African Rift, yet sparse, clustered observations limit reliable national inference. This study developed a leakage-controlled framework to classify groundwater fluoride exceedance above the WHO guideline of 1.5 mg/L, separating well-level prediction from national spatial transferability. A dataset of 1567 observations, including 567 exceedances, was evaluated using five-fold buffered spatial cross-validation with a 75-km train-test exclusion radius. The hydrochemical Extra Trees model (HYD_ET) achieved the strongest well-level performance, with Macro-F1 = 0.830, balanced accuracy = 0.817, MCC = 0.671, ROC-AUC = 0.910, PR-AUC = 0.859, and Brier score = 0.116. In contrast, the environmental-only Random Forest showed negligible geographic transferability (Macro-F1 = 0.495; balanced accuracy = 0.501), while Gaussian indicator kriging achieved Macro-F1 = 0.594. Applicability-domain analysis identified 8.23% of observations outside the model domain. Because no continuous model demonstrated adequate transferability, no wall-to-wall national F⁻-risk surface was retained. Instead, 48.43%, 18.84%, and 32.73% of 10-km grid cells had stronger, moderate, and weaker monitoring support, respectively. This evaluation spans heterogeneous hydrogeological and environmental settings under uneven monitoring across Ethiopia. The results show that strong well-level prediction does not imply reliable national extrapolation and supports groundwater monitoring and data collection.