Understanding the determinants of treated bed net use in Ethiopia: A machine learning classification approach using PMA Ethiopia 2023 survey data

Abstract

Introduction: Malaria remains a significant public health challenge in Ethiopia, with over 7.3 million cases and 1,157 deaths reported between January 1 and October 20, 2024. Despite extensive distribution campaigns, 35% of insecticide-treated nets (ITNs) remain underutilized, hindering malaria control efforts. Traditional statistical approaches have identified socioeconomic and demographic factors as predictors of ITN use, but often fail to capture complex, nonlinear interactions. This study applies machine learning to identify non-apparent factors of ITN utilization and investigates its performance in prediction as compared to traditional logistic regression.

Methods: This study applied ML models, including Random Forest, XGBoost, and Gradient Boosting, to predict ITN utilization using the 2023 Performance Monitoring for Action (PMA) Ethiopia dataset, a nationally representative survey of 9,763 households. The dataset included 18 variables: region, household size, wealth quintile, and housing conditions. Model performance was evaluated using accuracy, precision, recall, F1-score, and AUC-ROC. The values of SHAP (Shapley Additive Explanations) were used to interpret feature importance and interaction effects.

Results: Random Forest and XGBoost outperformed traditional logistic regression, achieving AUC scores of 0.89(0.91 after optimization) and 0.88, respectively. Key determinants of ITN utilization included geographic region, household size, wealth quintile, and maternal education. Nonlinear interactions, such as the moderating effect of maternal education on income-related barriers, were identified. Regional disparities were evident, with Amhara and Oromia showing higher ITN Utilization compared to urban areas like Harari and Dire Dawa. Middle-income households exhibited the highest ITN usage (23.7%), challenging the assumption of linear wealth gradients.

Conclusion: This study demonstrates the superiority of machine learning (ML) models in capturing complex, nonlinear determinants of ITN utilization, providing actionable insights for targeted malaria prevention strategies. Findings underscore the need for region-specific interventions, integration of ITN distribution with educational and economic empowerment programs, and synergies with environmental health improvements. The study highlights the potential of ML to enhance precision in public health in resource-limited settings, contributing to Ethiopia's National Malaria Elimination Roadmap and global malaria eradication efforts.

Fig 1. Distribution of treated bed net usage.

Fig 2. VIF score of independent variables.

Fig 3. Distribution of treated bed nets by region.

Fig 4. Distribution of treated bed nets by wealth quintile.

Fig 5. AUC-ROC curve comparison of trained models.

Fig 6. AUC-ROC result of the random forest model after hyperparameter tuning using a grid search.

Fig 7. Confusion matrix of the random forest model.

Fig 8. Top 10 feature importance for best performer random forest model.

Fig 9. SHAP interaction value.