Robust Machine Learning Methods for Prediction of Childhood Anemia – A Case of the Empowered Action Group States of India

Anemia is a major undernutrition concern in developing countries. Anemia in early childhood leads to lower immunity and diminished cognitive development and is one of the major causes of early childhood mortality. In India, the major burden of anemia is seen in the Empowered Action Group (EAG) states. Concerted efforts are needed to reduce the burden of anemia. This study uses machine learning (ML) models to predict anemia among children aged six to fifty-nine months using data from the fifth round of the Indian Demographic and Health Survey (DHS), also known as the National Family Health Survey – 5 (NFHS – 5) in the EAG states. The dataset had 85,189 rows. The random oversampling method was used to balance the dataset as there was a class-imbalance issue. Four ML models, namely conditional inference (CI) tree, random forest (RF), extreme gradient boosting (XGB), and k-nearest neighbors (KNN), were developed for prediction. The models were compared based on metrics such as accuracy, sensitivity, specificity, precision, and F1 score. The RF model had the best overall accuracy of 64%. The RF and XGB models had the best sensitivity of 0.75 and 0.7, respectively. The CI tree model had the highest specificity of 0.59. The RF and XGB models had the best F1 scores of 0.74 and 0.72, respectively. The RF model pointed out that the mother's nutritional status is the most important factor in predicting childhood anemia. Children were more likely to be anemic if their mothers had low Body Mass Index (BMI). This study contributes to the body of literature using ML techniques to study anemia in children.