Predicting Early Childhood Stunting and Associated Factors in Kenya Using Machine Learning

Abstract

Background: Stunting in the first two years of life remains a major public health concern in Kenya, affecting about one in five children and leading to long-term consequences on cognitive development and economic productivity. Traditional, largely reactive approaches that rely on linear statistical models have a limited ability to capture the complex, non‑linear determinants of stunting. Given that stunting arises from an interplay of immediate, underlying, and structural determinants, machine learning (ML) methods offer a promising alternative. ML techniques can leverage large, heterogeneous datasets to uncover hidden patterns and non‑linear interactions among associated factors, potentially enabling earlier and more accurate identification of children with higher probabilities of stunting. Objective of the study: This study applied supervised machine learning techniques to predict early childhood stunting among children aged 0-23 months and identify associated factors using nationally representative data from the 2022 Kenya Demographic and Health Survey. Methods: Three models—logistic regression (LR), balanced random forest (BRF), and categorical boosting (CatBoost)—were developed and evaluated using metrics aligned with programmatic priorities, including recall, area under the receiver operating characteristic curve (AUC), F1-score, and Brier Score. To address class imbalance, balanced class weights were applied, and models were trained at a threshold of 0.4 to optimize sensitivity. Results: CatBoost achieved the highest overall performance (AUC = 70.8%, recall = 85.8%), while LR remained highly competitive (AUC = 71.5%, Recall = 77.2%), outperforming BRF. Across all three models, consistent predictors included the child’s age, sex, birth size, twin status, maternal stature, household wealth, and geographic location, reflecting the multifactorial nature of stunting and aligning with existing literature. Conclusions: Simpler models, when combined with domain-informed feature engineering, could perform comparably to advanced algorithms while supporting transparency, which is essential for policy-relevant decision-making. The findings also provide methodological support for integrating predictive modelling into nutrition research and highlight opportunities to improve stunting prevention strategies.