Peer-reviewed article
Predicting zero-dose vaccination status in 27 sub-Saharan African countries: a machine learning approach
Beruk Berhanu Desalegn, Bewuketu Terefe ·
Summary
Written by AI, not yet AICS's view · claude-sonnet-5This study used machine learning to predict which children in sub-Saharan Africa are likely to receive zero vaccine doses, using pooled Demographic and Health Survey data from 27 countries collected between 2016 and 2024. Researchers tested seven different machine learning models and found that one called LightGBM performed best at identifying children at risk. Key factors linked to zero-dose status included where a mother gave birth, whether she attended antenatal care, her tetanus vaccination status, household wealth, and maternal education. The authors suggest this approach could help health systems identify at-risk children early so they can be reached with vaccination services.
Key points
- LightGBM achieved 80% accuracy, 91% sensitivity, and an AUROC of 0.78, the best among seven models tested
- Top predictors of zero-dose status were place of delivery, antenatal care attendance, maternal tetanus vaccination, household wealth, and maternal education
- Machine learning, particularly LightGBM and XGBoost, was found to be moderately effective at predicting zero-dose vaccination risk
- The model is proposed as a tool for early identification of children missing vaccinations, to support targeted intervention strategies
- Future work aims to scale the models using larger, routine health datasets for practical use
Limitations
The study is a predictive modelling exercise based on survey data and does not test or evaluate any actual intervention to increase vaccination coverage; it cannot show whether identifying at-risk children actually leads to improved vaccination rates.
Abstract
OBJECTIVE: To develop a machine learning (ML)-based predictive model for assessing the risk of zero-dose children using Demographic Health Survey data in sub-Saharan Africa. METHODS: This study analysed pooled Demographic and Health Survey data from 27 countries in sub-Saharan Africa collected between 2016 and 2024. Data preprocessing included imputation, balancing of unequal classes and systematic feature selection. Seven ML models were trained and evaluated using performance metrics such as accuracy, recall and F1-score. Feature importance was interpreted using SHapley Additive exPlanations (SHAP)-based analysis. RESULTS: Among the seven models evaluated, LightGBM demonstrated the best overall performance, achieving the highest accuracy (80%), sensitivity (91%), and area under the receiver operating characteristic curve (AUROC = 0.78). SHAP analysis identified place of delivery, antenatal care attendance, maternal tetanus vaccination, household wealth and maternal education as top predictors. DISCUSSION: Our findings suggest that machine learning, particularly LightGBM and XGBoost, is moderately effective in predicting the risk of zero-dose vaccination among children. This approach facilitates the easy identification of children most at risk of missing life-saving vaccinations. CONCLUSION: The LightGBM model predict children at zero-dose risk, enabling early identification and supporting intervention strategies. Future studies will scale these models using larger, routine datasets for practical implementation.