Predicting Tungiasis Hotspots in Western Kenya Using Machine Learning and Explainable Artificial Intelligence

Abstract

Tungiasis, a parasitic skin disease affecting millions across sub-Saharan Africa, remains one of the most neglected tropical diseases despite its substantial morbidity and socioeconomic burden. However, effective tools for identifying high risk populations and guiding targeted interventions remain critically limited. This study developed and evaluated an interpretable machine learning framework to predict household-level tungiasis risk in Western Kenya and identify the factors driving disease occurrence. Household level data from 5,876 households across five counties were used to train and validate six supervised machine learning algorithms. Model performance was evaluated using accuracy, precision, recall, F1 score, and the area under the receiver operating characteristic curve (AUC ROC), with 95% confidence intervals (CIs) estimated for all performance metrics. Explainable artificial intelligence techniques; SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) were applied to interpret model predictions. Random Forest achieved the highest point estimates for predictive performance (accuracy = 0.594, 95% CI: 0.559, 0.627; AUC ROC = 0.611, 95% CI: 0.574, 0.647), followed closely by Logistic Regression (accuracy = 0.578, 95% CI: 0.546, 0.609; AUC ROC = 0.608, 95% CI: 0.571, 0.643), with substantial overlap in confidence intervals across the leading models. Although overall predictive performance was moderate, considerable within county heterogeneity was observed (standard deviations: 0.122, 0.134), indicating that tungiasis risk is highly localized rather than uniformly distributed. Household income, rainfall, elevation, humidity, soil moisture, housing quality (particularly earthen floors), and footwear behavior were consistently identified as the strongest predictors of tungiasis risk. By integrating machine learning with explainable artificial intelligence, this study demonstrates that complex environmental, socioeconomic, and behavioral determinants of tungiasis risk can be identified and interpreted at both the population and household levels, providing actionable insights for targeted surveillance. While further validation and calibration are needed before operational deployment, this study provides a reproducible and interpretable framework that can support evidence-based surveillance and targeted control strategies in tungiasis-endemic settings.