(LB03-23-25) Determining Food Groups Associated With Height in Adolescence Using a Novel Supervised Learning Approach

Objectives: Food Frequency Questionnaires (FFQs) are essential for assessing dietary intake, but evaluating all foods in relation to health outcomes in one statistical model presents analytic challenges. This study applies a novel statistical method to identify food groups strongly associated with adolescent height. By identifying key food groups, we derive adaptive summary scores that enhance statistical power and improve clinical interpretation.

Methods: The study included 396 participants (mean age: 13.89 ± 1.94 years) in the Early Life Exposure in Mexico to Environmental Toxicants (ELEMENT) cohort study. Dietary intake in the past 7 days was assessed using an FFQ adapted for the Mexican population, categorizing >100 food items into 33 groups with consumption frequency scores with corresponding calorie estimates. We implemented a best-subset selection algorithm based on mixed integer optimization (MIO) to identify associated food groups with height while balancing model complexity and interpretability. This method refines the FFQ summary metric for dietary assessment. Estimation uncertainty was assessed via noise perturbation, and the methodology was validated through simulations. The model was then applied to explore associations between dietary intake and height.

Results: Five foods-Milk, Pork, Processed Meat, Sweetened Milk, and Unsaturated Spread-were positively associated with greater height (β=0.004cm per unit increase in food group summary score, p=0.003), adjusting for age, sex, weight, socioeconomic status, and maternal education. Additionally, greater height was significantly associated with being older (β=3.682cm/year, p< 0.001), heavier weight (β=4.158cm/kg, p< 0.001), and male (β=2.179, p< 0.001) adolescents. The summary score from the selected food groups was statistically significant, and the resulting model had a higher adjusted R² (0.628) compared to standard linear regression without performing selection (0.620).

Conclusions: The MIO method enhances feature selection, improving the interpretability of dietary influences on adolescent height while maintaining predictive power. Our findings highlight key food groups associated with growth and demonstrate the potential of supervised learning in nutritional epidemiology.

Funding Sources: RD83543601, P20 ES018171, P01 ES02284401, P30 ES017885, U24 ES028502.