671 - Exploratory identification of features associated with child development among Lebanese children aged two to five: A nationally representative analysis using ensemble machine learning

Publication Number: 3649.671

Lama Charafeddine, American University of Beirut, Beirut, Beyrouth, Lebanon; Alya Al Sager, Harvard University, Boston, MA, United States; Junita M. Henry, Harvard University, Boston, MA, United States; Amirhossein Yarparvar, unicef, kabul, Kabol, Afghanistan; Carla El-Mallah, GroundWork, Fläsch, Graubunden, Switzerland; Joelle Najjar, UNICEF, Achrafiyeh, Beyrouth, Lebanon; Pamela Zgheib, AdventHealth for Children, beyrouth, Beyrouth, Lebanon; Janneke H.. Blomberg, UNICEF, Nykoebing Falster, Syddanmark, Denmark; Aisha K. Yousafzai, Harvard TH Chan School of Public Health, Harvard University, Boston, MA, United States

Background: Understanding the determinants of early childhood development (ECD) in crisis-affected settings remains a central challenge for global health and social policy.

Objective: To identify correlates of developmental outcomes applying machine-learning models.

Design/Methods: Using nationally representative data from the 2023 Lebanon Integrated Micronutrient and Anthropometric Survey, we applied two ensemble machine-learning models, eXtreme Gradient Boosting (XGBoost) and Bayesian Additive Regression Trees (BART), to identify correlates of developmental outcomes among 812 children aged 24–59 months. Models incorporated the survey’s complex design and population weights and were multiply imputed to address missing data. The study received IRB approval and participants gave their written consent.

Results: Both models achieved comparable weighted predictive accuracy yet emphasized different domains. XGBoost prioritized socioeconomic position, nutrition, and access to services, while BART highlighted caregiving and household processes, including non-violent discipline, shared decision-making, and recent child health. Structural deprivation and caregiving quality thus emerged as complementary dimensions of developmental vulnerability.

Conclusion(s): These results demonstrate that ensemble methods can uncover both robust and model-contingent associations, revealing how developmental risk operates across structural and relational domains. The approach provides a transparent, reproducible framework for integrating interpretable machine learning with survey-based ECD measurement in fragile contexts. While the findings are correlational, they emphasize the multidimensional nature of early development and the need for integrated policies that address material, nutritional, and caregiving environments simultaneously.