Predicting health insurance coverage among women in Sierra Leone using machine learning and explainable artificial intelligence applied to demographic and health survey data

Health insurance coverage remains critically low in Sierra Leone, creating barriers to achieving universal health coverage (UHC) and leaving women particularly vulnerable to catastrophic out-of-pocket health expenditures. Understanding which women are most at risk of being uninsured and the specific combination of factors perpetuating their exclusion is essential for designing effective, targeted interventions. This study aimed to develop and evaluate machine learning models to predict health insurance coverage among women in Sierra Leone, identify the most influential sociodemographic and economic predictors, and translate model outputs into actionable, precision-targeted policy recommendations for advancing UHC. Data from the 2019 Sierra Leone Demographic and Health Survey (SLDHS), a nationally representative dataset of 15,574 women, were used. Six classifiers were compared: Logistic Regression, Support Vector Machine, k-Nearest Neighbours, Random Forest, Gradient Boosting, and XGBoost. Cost-sensitive learning strategies were applied to address pronounced class imbalance, where insured women constituted only 3.5% of the sample. To enhance interpretability and support policy translation, Explainable AI (XAI) frameworks, SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations), were applied to the best-performing interpretable model. Gradient Boosting achieved the highest overall performance, with an AUROC of 0.8857 and a minority class recall of 80.6%. XGBoost performed competitively with an AUROC of 0.869 and was selected for interpretability analysis due to its established compatibility with tree-based explanation workflows. Global SHAP analysis identified region, wealth index, and radio listenership as the three most influential predictors of insurance coverage, with decision-making autonomy variables demonstrating meaningful but secondary contributions. Instance-level waterfall plots and LIME explanations revealed substantial heterogeneity in insurance pathways, distinguishing structurally advantaged, compensatory, and compound-disadvantage profiles among women. Ensemble machine learning models, particularly Gradient Boosting and XGBoost, demonstrate strong predictive utility for identifying uninsured women in low-coverage, resource-limited settings. The integration of SHAP and LIME frameworks bridges the gap between algorithmic performance and actionable policy insight, revealing that insurance coverage is strongly associated with geographic location, economic status, and media exposure. These findings provide an evidence-based framework for precision public health strategies that simultaneously address geographic disparities, economic barriers, and women’s empowerment constraints in Sierra Leone.