Nationwide Spatiotemporal Dynamics and Machine Learning Prediction of Anemia Among Women in Lesotho

BackgroundDespite substantial efforts, anemia continues to pose a significant public health challenge, disproportionately affecting women of reproductive age. In Lesotho, insufficient data exist on its geographic distribution and key determinants using advanced techniques. This study seeks to address this gap, contributing to the development of interventions and thereby accelerating the pathway to reducing its burden. MethodsThis study used LDHS 2023/24 data on women of reproductive age (15-49 years), collected from 27 November 2023 to 29 February 2024 via a two-stage stratified sampling design. Spatial analysis included Global Morans I for autocorrelation, Getis-Ord Gi* for hotspot detection, Bayesian empirical kriging for unsampled locations, and Bernoulli-based purely spatial analysis to identify significant anemia clusters. To address class imbalance, SMOTE was applied, and feature selection was conducted using the Boruta algorithm. Multiple machine learning models were trained to predict anemia, with performance evaluated using accuracy, sensitivity, specificity, precision, F1-score, and ROC-AUC, and interpretability enhanced via SHAP method. Additionally, association-rule mining with the Apriori algorithm explored potential interactions among variables. ResultsThe prevalence of anemia was 23.7%. Anemia exhibited significant spatial dependency (Morans I = 0.248, p < 0.001, Z-score = 7.774), indicating a non-random distribution. Hotspot analysis identified higher clustering in northwestern Lesotho (Leribe) and southwestern districts (Mohales Hoek and Quthing), while lower clustering was observed in Qachas Nek. The primary cluster, spanning western and central districts, represented the most pronounced high-risk area, with a relative risk of 2.34 (p < 0.001). The Gradient Boosting Machine demonstrated strong predictive performance (AUC = 87.1%), with balanced sensitivity (72.5%) and specificity (80.3%). Low education, number of children under five, poor wealth, unimproved toilet facility, occupation, urban residence, and age (35-49 years) emerged as the most influential determinants of anemia. ConclusionsAnemia affects nearly one in four women in Lesotho, resulting from a complex interplay of socioeconomic, demographic, and environmental factors. Machine learning analysis, particularly the Gradient Boosting Machine, demonstrated strong predictive performance, highlighting its usefulness for predicting anemia. Effective interventions should combine geographically targeted programs in hotspot areas with broader strategies addressing womens education, economic empowerment, sanitation, and nutrition.