Machine learning-based prediction of limited access to drinking water services among households using improved water sources in East Africa: evidence from recent DHS data
Globally, more than 2.1 billion people lack access to improved drinking water services. Limited access to drinking water services has become one of the most critical challenges in developing countries, especially in East Africa. Despite this huge public health burden associated with limited water access, policymakers have given less attention to improving water service accessibility, and less research has been done among those households using improved water sources. Identifying the key predictors of limited access to drinking water can help to design targeted interventions. Therefore, this study aimed to develop and evaluate machine learning models for predicting limited access to drinking water services among households using improved water sources in East Africa. This cross-sectional study employed machine learning models using the most recent DHS data from 146,120 households with improved drinking water sources. Data preprocessing involved handling missing data using k-Nearest Neighbors and mode imputation, encoding categorical features, and hybrid feature selection. Class-weighting, SMOTE, and SMOTEENN were compared to address class imbalance. Logistic Regression, Random Forest, Decision Tree, K-Nearest Neighbors, CatBoost, Support Vector Machine, Naïve Bayes, ANN, and XGB were trained on 80% of the data, with 20% reserved for testing. Model performance was evaluated using accuracy, precision, recall, F1-score, AP, and ROC-AUC. SHAP was used to interpret feature contributions to predictions, while the Apriori algorithm was used to identify co-occurring patterns. The study found that in East Africa, among households using improved water sources, 22.1% (95% CI: 21.4%–22.8%) had limited access to drinking water services. The CatBoost model demonstrated the best predictive performance (AUC = 80.8% and AP = 49.8%). SHAP analysis identified responsibility for water collection by a male child under 15 years, household wealth status, community-level electricity access, country, and place of residence as the strongest contributors to model predictions. Apriori association rule mining identified household poverty, low community education, and poor water treatment practices in the community as the strongest co-occurring predictors of limited access to drinking water services. CatBoost with Bayesian optimization showed the strongest performance and was selected as the final model for interpreting predictors. These findings provide valuable insights for targeted interventions, including improving the proximity and availability of drinking water services, reducing the burden of water collection on children and women, addressing socioeconomic and infrastructure-related barriers, and strengthening community-level capacity to improve equitable access to drinking water services.
Authors
- Gete Berihun (ORCID: https://orcid.org/0000-0002-5511-2072)
- Abraham Keffale Mengistu (ORCID: https://orcid.org/0009-0000-0014-2242)
- Endeshaw Zemenu Enatu
- Erimias Getsh Teshome
- Kirubel Genet Nibretie
Institutions
- Wollo University (ET)
- Debre Markos University (ET)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-13
- DOI
- https://doi.org/10.1038/s41598-026-78378-5
- Primary Topic
- Water Resource Management and Quality
- Type
- article
- Field-Weighted Citation Impact
- 0.00