Interpretable machine-learning prediction of household risk of trachoma and spatial hotspot mapping in Ethiopia

BACKGROUND: Active trachoma remains above elimination thresholds in Ethiopia, yet tools for fine-scale targeting are limited. We developed and internally validated an interpretable machine-learning model to predict household-level risk of trachomatous inflammation-follicular (TF) and mapped hotspots to guide SAFE implementation. METHODS: A community-based cross-sectional survey was conducted in Dessie Zuria District (March-May 2024) among households with children aged 1-9 y. TF was graded using the amended WHO simplified system. Behavioral, environmental and sociodemographic predictors were collected using WHO-aligned tools. A survey-weighted ridge-penalized logistic regression with nested cross-validation was trained, with missing data imputed within folds. Model performance (AUC, calibration, Brier score) and decision-curve analysis were evaluated using kebele-cluster bootstrapping. Spatial clustering was assessed using Moran's I and false discovery rateadjusted Getis-Ord Gi*. RESULTS: Among 612 households (928 children), 9.8% of households and 9.1% of children were TF-positive. Key predictors included unclean child face, absence of latrine, water-collection time >30 minutes, flies on the face and low hygiene score. The model showed strong discrimination (AUC = 0.84), good calibration (slope 0.98) and accuracy (Brier 0.067). CONCLUSIONS: The model identifies high-risk households and spatial hotspots, supporting targeted F/E interventions to accelerate trachoma elimination.