The performance of machine learning algorithms in predicting under-five anaemia in Tanzania

Anaemia in children under five remains a major public health concern, contributing significantly to global child morbidity and mortality. With the rise of machine learning (ML), novel opportunities exist to model and predict anaemia more effectively. This study aimed to evaluate and compare the performance of five ML algorithms in predicting anaemia among under-five children in Tanzania. We conducted a secondary data analysis using the 2017 Tanzania Malaria Indicator Survey (TMICS). The dataset (n = 5906) was randomly split into training (70%) and test (30%) sets. Five ML algorithms, Linear Discriminant Analysis (LDA), Logistic Regression, Random Forest (RF), Least Absolute Shrinkage and Selection Operator (LASSO), and Ridge Regression were trained using fivefold cross-validation and evaluated on the test set. ROC-AUC, accuracy, precision, recall, and F1-Score metrics were used to assess model performance. Largely sociodemographic variables were used as features, and analyses were performed in both R and Python. Among the children, 51% were male, 64% were over two years, and 74% resided in rural areas. Anaemia prevalence was 59%. Prediction accuracy ranged from 61 to 62% across models. The precision, recall, and F1-Score metrics were similar across models in the training and test sets, with the exception of the random forest which showed signs of overfitting. AUC values were ~ 67% for all models, except the random forest, which showed AUC value of 60%. The models largely showed relatively weak performance and discrimination power. Although the model metrics do not suggest clinical utility, the study demonstrates a proof of concept in the potential of ML techniques in modelling childhood anaemia using population health data.