Abstract
Byline: Madison N LeCroy, NYU Grossman Sch of Medicine, New York, NY; Ryung S Kim, Albert Einstein College of Medicine, Bronx, NY; David B Hanna, Albert Einstein College of Medicine, Bronx, NY; Krista M Perreira, Univ of North Carolina Sch of Medicine, Chapel Hill, NC; Linda C Gallo, San Diego State Univ, Chula Vista, CA; Maria M Llabre, Univ of Miami, Coral Gables, FL; Linda Van Horn, Northwestern Univ Feinberg Sch of Medicine, Chicago, IL; Martha L Daviglus, Univ of Illinois Chicago, Chicago, IL; Gregory A Talavera, San Diego State Univ, Chula Vista, CA; Daniela Sotres-Alvarez, Univ of North Carolina at Chapel Hill, Chapel Hill, NC; Carmen R Isasi, Albert Einstein College of Medicine, Bronx, NY Introduction: Among 2-19-year-olds in the United States (US), 26.2% of Hispanic/Latino youth vs. 16.6% of non-Hispanic White youth experience obesity (body mass index [BMI] percentile [greater than or equal] age- and sex-specific 95th BMI percentile). While health behaviors are important, psychological and sociocultural measures vary across racial/ethnic groups and may underpin obesity disparities. Machine learning is one statistical approach that can be used to identify determinants of obesity. However, few studies have applied these methods to childhood obesity research, with most studies only examining traditional risk factors and creating a single model across all racial/ethnic groups. Our objective was to identify key predictors of BMI percentile in Hispanic/Latino youth to help design childhood obesity interventions that reduce health disparities. Hypothesis: We hypothesized that a BMI percentile prediction model developed for Hispanic/Latino youth would identify both traditional and novel risk factors as important determinants. Methods: Hispanic/Latino 8-16-year-olds from the 4 US sites of the Hispanic Community Children's Health Study/Study of Latino Youth (SOL Youth) were examined (n=1,466). BMI percentiles were determined via measured height and weight and CDC growth charts. A supervised machine learning approach, Least Absolute Shrinkage and Selection Operator (LASSO) regression, was used with BMI percentile as the outcome. There were 98 predictor variables examined spanning demographics; health behaviors; and environmental, psychological, and sociocultural measures. LASSO-selected variables were entered into a multivariable linear regression model to obtain effect estimates. P-values were adjusted for both multiple testing and the variable selection process and assessed with ð<0.025. Models incorporated survey weights, and missing data were imputed. Results: A LASSO model with 30 variables yielded the optimum mean squared error (MSE; R2=0.37), but a 12-variable solution was selected based on MSE and parsimony. Four associations were statistically significant. Perception of being larger than the "ideal" body weight (ð=8.75 [95% CI: 8.21, 12.13]), reporting disordered eating (ð=12.61 [95% CI: 10.64, 14.91]), and having a parent with obesity (ð=4.90 [95% CI: 3.49, 17.75]) were associated with a higher BMI percentile. Spending >$5 weekly on snacks/beverages/fast food (ð=-4.62 [95% CI: -19.18, -2.09]) was associated with a lower BMI percentile. Conclusions: Psychological and parental factors predicted higher BMI percentile and greater money spent on snacks/beverages/fast food predicted lower BMI percentile among Hispanic/Latino youth in the US. Addressing Hispanic/Latino youth's relationships with food, body weight, and parents may be important in obesity interventions. Longitudinal research is needed to clarify directionality and replicate novel findings.