JMIR Publications

ABSTRACT

Background:

Background:

Anemia during the postnatal period remains a severe public health crisis in resource limited settings such as Ethiopia, contributing significantly to maternal morbidity and mortality. Machine learning (ML) paradigms offer robust methodologies for early detection, risk stratification and data driven clinical intervention. This study develops and validates highly interpretable machine learning classification models to predict anemia severity among postnatal women using large scale nationally representative data from the recent 2024–25 Ethiopian Demographic and Health Survey (EDHS).

Objective:

Develops and validates highly interpretable machine learning classification models to predict anemia severity among postnatal women using large scale nationally representative data from the recent 2024–25 Ethiopian Demographic and Health Survey (EDHS)

Methods:

Methods:

Utilizing an expanded multistage stratified sample of 45280 postnatal records derived from the 2024/25 EDHS combined with institutional health registries, we evaluated six distinct classification algorithms: Multi-layer Perceptron Neural Network (MLP-NN), Extreme Gradient Boosting (XGBoost), Gaussian Naive Bayes (GNB), Random Forest (RF), Decision Tree (DT) and K-Nearest Neighbors (KNN). Features were systematically selected using Mutual Information (MI) and F-scores. Performance metrics included accuracy, sensitivity, specificity, precision, F1-score, Mean Absolute Error (MAE) and rigorous p-value hypothesis testing.

Results:

Results:

The Random Forest classifier demonstrated superior predictive performance across all clinical categories achieving an exceptional overall accuracy of 97.4%, a precision of 93.2%, recall of 93.5% and an F1 score of 93.3% (MAE = 0.041). Decision Tree and XGBoost models also performed robustly with accuracies of 96.2% and 95.8% respectively. Postnatal determinants including Body Mass Index (BMI), frequency of antenatal care (ANC) visits, household wealth index and regional stratification emerged as the most prominent predictors of anemia severity.

Conclusions:

Conclusions:

Interpretable machine learning models specifically Random Forest, can accurately predict anemia severity among postnatal women at a national scale. Integrating these computational algorithms into electronic health record workflows can empower healthcare providers in Ethiopia to prioritize high risk mothers for targeted iron supplementation and nutritional therapies.