Development of a Deep Learning Model for Predicting Obesity Using Health Behavior Data of Elementary School Students

Abstract

Background: Childhood obesity poses serious long-term health risks and is a growing global concern. In South Korea, national health surveys collect behavioral and physical data from elementary students, but the large number of questionnaire items can burden young respondents and reduce accuracy. Thus, simplified models with high predictive power are needed.

Methods: We analyzed data from over 250,000 elementary students collected by the Korean Ministry of Education (2015–2022). Using the Rohrer Index as the outcome variable, key predictors were selected via Lasso and Elastic Net regression. Categorical variables were reduced using Multiple Correspondence Analysis (MCA), and a deep learning model (NECTOR) combining MLP and self-attention was developed.

Results: NECTOR achieved high predictive performance with R² scores of 0.994 (boys) and 0.996 (girls), and low mean squared errors of 3.072 and 1.841, respectively. It outperformed baseline models using the same inputs.

Conclusion: A small set of core health indicators can effectively predict the Rohrer Index. The proposed model enables efficient and reliable obesity screening in school settings, supporting early intervention efforts.