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

Area of Science:

• Pediatric Health
• Biostatistics
• Machine Learning

Background:

• Childhood obesity is a significant global health issue with long-term consequences.
• South Korean national surveys collect extensive student data, but lengthy questionnaires may compromise accuracy and respondent burden.
• There is a need for simplified, high-predictive models for efficient childhood obesity assessment.

Purpose of the Study:

• To develop a simplified, accurate model for predicting childhood obesity using a limited set of health indicators.
• To evaluate the predictive performance of a novel deep learning model (NECTOR) for obesity screening.

Main Methods:

• Analysis of data from over 250,000 South Korean elementary students (2015-2022).
• Selection of key predictors using Lasso and Elastic Net regression for the Rohrer Index.
• Development of the NECTOR deep learning model, integrating MLP and self-attention, after reducing categorical variables with Multiple Correspondence Analysis (MCA).

Main Results:

• NECTOR demonstrated high predictive accuracy with R-squared values of 0.994 for boys and 0.996 for girls.
• The model achieved low mean squared errors (3.072 for boys, 1.841 for girls).
• NECTOR significantly outperformed baseline models utilizing the same input variables.

Conclusions:

• A concise set of core health indicators is sufficient for effective Rohrer Index prediction.
• The NECTOR model provides a feasible solution for efficient and reliable obesity screening in school environments.
• This approach supports timely interventions for childhood obesity.