Predicting Body Composition In The U.S. Population Using Machine Learning Models

Research using machine learning approaches to predict body fat percentage (BF%) often overlooks the complex survey design and missing data imputation. PURPOSE: This study aims to bridge this gap by examining how incorporating sampling weights and multivariate imputation by chained equations (MICE) affects random forest (RF) prediction accuracy. METHODS: This study utilized the 2005-2006 National Health and Nutrition Examination Survey (NHANES; 8-69 years old, observation = 6,469, 49.25% females) and included predictors such as demographics, anthropometrics, mental health, healthy lifestyle behaviors, and disability status. Four RF regression models were evaluated: (1) Without MICE or sampling weights (M1); (2) Without MICE but with sampling weights (M2); (3) With MICE but without sampling weights (M3); (4) With MICE and sampling weights (M4). The optimal model was defined by the lowest Mean Squared Error (MSE). Variable importance indexes were extracted to identify which predictors most significantly influenced the model’s predictions. Partial dependence plots (PDP) were used to visualize the relationship between BF% and different predictors. RESULTS: The RF regression model with MICE and sampling weights (M4) was optimal, with the lowest MSE of 8.984. This was followed closely by M3 (MSE = 9.015), then M1 (MSE = 9.065) and M2 (MSE = 9.140). In the optimal model, triceps skinfold was the most important predictor, followed by sex, subscapular skinfold, body mass index (BMI), waist circumference (WC), height, weight, age, and physical activity metrics (counts and steps). The PDP indicated that higher BF% was associated with thicker triceps and subscapular skinfolds, females, higher BMI, larger WC, and decreased physical activity. An interesting finding showed that as children and adolescents aged, their BF% decreased; However, as adults aged, their BF% increased. CONCLUSIONS: The results suggest that using sampling weights and MICE improves the accuracy of RF models in predicting BF%. The study successfully developed a model with a low MSE, effectively predicting BF% across a broad age range.