[Paper Review] Predicting Anemia Among Under-Five Children in Nepal Using Machine Learning and Deep Learning

Deepak Bastola, Pitambar Acharya+3 more|arXiv (Cornell University)|Feb 1, 2026

Iron Metabolism and Disorders0 citations

TL;DR

This study benchmarks 10 ML/DL models to predict anemia in Nepalese children (6-59 months) using NDHS 2022 data, with 13 sociodemographic features and SMOTE handling of class imbalance.

ABSTRACT

Childhood anemia remains a major public health challenge in Nepal and is associated with impaired growth, cognition, and increased morbidity. Using World Health Organization hemoglobin thresholds, we defined anemia status for children aged 6-59 months and formulated a binary classification task by grouping all anemia severities as \emph{anemic} versus \emph{not anemic}. We analyzed Nepal Demographic and Health Survey (NDHS 2022) microdata comprising 1,855 children and initially considered 48 candidate features spanning demographic, socioeconomic, maternal, and child health characteristics. To obtain a stable and substantiated feature set, we applied four features selection techniques (Chi-square, mutual information, point-biserial correlation, and Boruta) and prioritized features supported by multi-method consensus. Five features: child age, recent fever, household size, maternal anemia, and parasite deworming were consistently selected by all methods, while amenorrhea, ethnicity indicators, and provinces were frequently retained. We then compared eight traditional machine learning classifiers (LR, KNN, DT, RF, XGBoost, SVM, NB, LDA) with two deep learning models (DNN and TabNet) using standard evaluation metrics, emphasizing F1-score and recall due to class imbalance. Among all models, logistic regression attained the best recall (0.701) and the highest F1-score (0.649), while DNN achieved the highest accuracy (0.709), and SVM yielded the strongest discrimination with the highest AUC (0.736). Overall, the results indicate that both machine learning and deep learning models can provide competitive anemia prediction and the interpretable features such as child age, infection proxy, maternal anemia, and deworming history are central for risk stratification and public health screening in Nepal.

Motivation & Objective

• Assess whether machine learning and deep learning models can predict anemia status (any severity) among children aged 6-59 months in Nepal using NDHS 2022 data.
• Identify robust feature subsets by applying multiple feature selection techniques and consensus among methods.
• Evaluate model performance with appropriate metrics under class imbalance, and determine practical screening potential for resource-limited settings.
• Examine the relative strengths of interpretable models (e.g., logistic regression) versus neural network approaches in this context.

Proposed method

• Preprocess NDHS 2022 data by removing entries with missing targets and high-missing variables; impute remaining gaps.
• Convert features to categorical representations and apply 80/20 train-test split with stratified sampling.
• Use four feature selection methods (Chi-square, Mutual Information, Point-Biserial, Boruta) and select features supported by multi-method consensus.
• Balance training data with SMOTE and integrate into the modeling pipeline to avoid data leakage.
• Train and optimize ten models (LR, KNN, DT, RF, XGBoost, SVM, NB, LDA, DNN, TabNet) using repeated stratified cross-validation and grid search.
• Evaluate models with accuracy, precision, recall, F1-score, average precision, AUC, and Cohen’s kappa; emphasize recall and F1 due to class imbalance.

Figure 1: Normalized features importance across methods

Experimental results

Research questions

• RQ1 Can ML/DL models predict anemia status in Nepalese children using sociodemographic features from NDHS 2022?
• RQ2 Which features are most consistently informative for anemia risk across multiple selection methods?
• RQ3 How do different ML/DL algorithms perform under SMOTE-balanced training in this setting?
• RQ4 Is logistic regression competitive with neural networks for this tabular, small-sample, highly imbalanced problem?
• RQ5 What implications do model performances have for non-invasive screening and public health screening programs in Nepal?

Key findings

• Logistic regression achieved the best recall (70.1%) and highest F1-score (64.9%), indicating strong ability to identify anemic cases.
• SVM provided the highest AUC (0.736), showing strong discrimination between anemic and non-anemic children.
• DNN achieved the highest accuracy (70.9%) and precision (66.4%), but had a lower recall (60.4%) than LR.
• Overall model performance was in the 60-70% range across metrics, reflecting the predictive limits of sociodemographic proxies versus direct clinical measures.
• Top predictive features included child age, fever, household size, maternal anemia, and parasite deworming, with several ethnicity and provincial indicators also informative.
• Models consistently benefited from SMOTE to address class imbalance, and cross-validation suggested robust generalization.

Figure 3: Train and test ROC curves