Related Experiment Video
Updated: Aug 14, 2026
A Point-of-Care Method with Integrated Decision Support Tool to Estimate Anemia at Population Level
Published on: January 19, 2024
Explainable machine learning for predicting childhood anemia in Sub-Saharan Africa using population-based DHS Data
Andualem Enyew Gedefaw1, Amanuel Worku2, Abraham Keffale Mengistu3
1Department of Health Informatics, Institute of Public Health, College of Medicine and Health Sciences, University of Gondar, Gondar, Ethiopia.
Insights
Childhood anemia prediction in Sub-Saharan Africa shows machine learning models offer moderate performance. While statistically significant improvements over traditional methods were observed, low sensitivity necessitates complementary screening strategies for effective intervention.
Area of Science:
- Public Health
- Computational Epidemiology
- Pediatrics
Background:
- Childhood anemia is a critical public health issue in Sub-Saharan Africa, impacting child development and survival.
- Prevalence exceeds 60% across 26 countries, demanding scalable prediction tools for targeted interventions.
Purpose of the Study:
- To evaluate the performance of various machine learning models in predicting childhood anemia using Demographic and Health Survey (DHS) data.
- To identify key predictors of childhood anemia for improved public health strategies.
Main Methods:
- Pooled DHS data from 26 Sub-Saharan African countries (110,251 children aged 6-59 months) were analyzed.
- Eight machine learning models were trained and evaluated using hyperparameter tuning and 5-fold cross-validation.
- Model performance was assessed via accuracy, precision, recall, F1-score, ROC-AUC, with SHAP for interpretability.
Main Results:
- The CatBoost model achieved the highest performance (ROC-AUC = 0.84), outperforming logistic regression significantly (p < 0.001).
- Models showed moderate discriminatory ability with limited sensitivity (recall ≈ 0.33-0.41).
- Key predictors included residence type, height-for-age z-score, country, and child age.
Conclusions:
- Machine learning models offer moderate to strong predictive performance for childhood anemia using DHS data.
- Despite statistical significance, low sensitivity restricts their use as standalone screening tools.
- Findings emphasize early nutrition, undernutrition reduction, and integrated predictive analytics for resource allocation in high-burden regions.
Abstract:
Childhood anemia remains a major public health challenge in Sub-Saharan Africa, adversely affecting physical growth, cognitive development, and child survival. The pooled prevalence across 26 countries exceeds 60%, underscoring the need for accurate and scalable prediction tools to support targeted interventions.This study used pooled Demographic and Health Survey (DHS) data (2016-2024) from 26 Sub-Saharan African countries, including 110,251 children aged 6-59 months. Multiple machine learning models (Logistic Regression, Decision Tree, Extra Trees, Random Forest, XGBoost, LightGBM, CatBoost, and MLP) were trained using hyperparameter tuning with 5-fold cross-validation. Model performance was evaluated using accuracy, precision, recall, F1-score, and ROC-AUC, with 95% confidence intervals estimated via bootstrapping. Model comparisons were conducted using the DeLong test, and SHAP was used for model interpretability.The CatBoost model demonstrated the best overall performance (ROC-AUC = 0.84, 95% CI: 0.840-0.848), followed closely by XGBoost and LightGBM. All machine learning models significantly outperformed logistic regression (p < 0.001, DeLong test), although the absolute improvement in discrimination was modest (ΔAUC ≈ 0.04). The models demonstrated moderate discriminatory ability, with relatively low to moderate sensitivity (recall ≈ 0.33-0.41) depending on the model and classification threshold. SHAP analysis identified residence type, height-for-age z-score, country, and child age as the most influential predictors.Machine learning models demonstrated moderate to strong discriminatory performance in predicting childhood anemia using DHS data. Although improvements over traditional models were statistically significant, the relatively low sensitivity limits their effectiveness as standalone screening tools. These findings highlight the importance of early childhood nutrition (particularly during 6-23 months), reduction of chronic undernutrition, and context-specific public health strategies. Integration of predictive analytics into national health systems may support risk stratification and resource allocation in high-burden settings. However, findings should be interpreted in light of the cross-sectional design and lack of external validation.
