Predicting BMI Percentile in Hispanic/Latino Youth Using a Machine Learning Approach: Findings From the Study of Latino Youth.

Abstract

The objective of this study is to use a machine learning approach to identify predictors of BMI percentile among Hispanic/Latino youth in the United States. Participants were Hispanic/Latino 8- to 16-year-olds from the cross-sectional Study of Latino Youth (SOL Youth; n = 1466). A supervised machine learning approach, LASSO regression, was used with BMI percentile as the outcome. A total of 102 predictor variables were examined spanning parent and child demographics; health behaviors; and psychological, sociocultural, and environmental measures. Mean age of participants was 12 years, 50% were female, and 44.2% were of Mexican heritage. A 36-variable LASSO model yielded the optimum mean squared error (R ² = 0.42), but a 10-variable solution was selected for parsimony. Six associations were significant. Dieting 1-4 or ≥ 5 times/year (β = 8.69 [95% CI: 10.25 to 14.52] or 10.86 [95% CI: 13.14 to 18.33], respectively) and having a parent of Dominican heritage (β = 3.48 [95% CI: 4.05 to 9.90]) or with obesity (β = 2.96 [95% CI: 2.99 to 6.85]) were associated with a higher BMI percentile. Perception of being smaller than the "ideal" body size (β = -1.65 [95% CI: -6.84 to -1.35]) and use of the food/activity parenting practice Control (β = -1.17 [95% CI: -3.63 to -1.69]) were associated with a lower BMI percentile. Family-based approaches and focusing on dieting and body image satisfaction may be important for weight management in Hispanic/Latino youth.