A Machine Learning Approach to Predicting Household Smoke Exposure Risk in Somalia: An Analysis With SHAP Explanations.

Introduction: Household air pollution (HAP) from solid fuel combustion is a major global public health issue with a particularly high burden in sub-Saharan Africa. In Somalia, the extent and predictors of household smoke exposure risk (SER) remain underexplored due to data scarcity and analytical limitations. This study applies machine learning (ML) models to identify and predict SER in Somali households using the first Somalia Demographic and Health Survey (SDHS) and interpretable artificial intelligence (AI) techniques. Methods: A nationally representative sample of 15 838 households from the 2020 SDHS was analyzed using multivariate logistic regression. The SER was defined based on the cooking fuel type and location. Six supervised ML models (Logistic Regression, K-Nearest Neighbors, Decision Tree, Support Vector Machine, Random Forest, Gradient Boosting) were trained using an 80/20 train-test split. The performance was evaluated using accuracy, precision, recall, F1-score, and AUROC. Feature importance was assessed using Gini, permutation, and SHAP (SHapley Additive explanation) values. Results: The prevalence of household smoking exposure was 70.0%. Place of residence, region, and wealth were the dominant predictors. Gradient Boosting outperformed other models (AUC = 81% [95% CI: 79.11%-82.17%], Conclusion: This is the first national study to apply machine learning to predict SER in Somalia, revealing urban and wealth-linked vulnerabilities that challenge conventional assumptions about poverty. These findings highlight the need for targeted clean cooking interventions in urban and peri-urban communities, alongside data innovations for real-time monitoring. ML-informed risk stratification may support more effective and equitable health policies in fragile states.