Prediction of Device-Derived Body Age Using Bioelectrical Impedance Analysis and Graded Treadmill Performance: An Interpretable Machine Learning Study
Wafaa Mohmoud Abdellatif Bekir, Fatma Hassan Abd Elbasset Mourgan, Manaf AlMatarAbstract
Background:
Body Age, a simplified indicator of health and fitness, is increasingly available as a value derived by commercial bioelectrical impedance analysis devices. The purpose of this study was to train and internally validate machine learning models to predict device-derived Body Age throughout a leakage-aware set of anthropometric, body-composition, metabolic, and graded treadmill-performance variables.
Methods:
This exploratory study used a secondary analysis of 55 de-identified body-composition baseline records and graded treadmill test records. To reduce target leakage, 20 independent predictor variables were selected after removing proprietary control parameters and composite evaluation outputs. Elastic Net regression, Support Vector Regression (SVR), and Gradient Boosting Regression (GBR) were developed and internally validated using nested cross-validation, with five outer folds for performance estimation and four inner folds for hyperparameter selection. Missing values were imputed with the median within each validation pipeline, and feature standardization was performed when applicable. Model performance was evaluated using mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (
Results:
Device-derived Body Age ranged from 15 to 30 years (19.35 ± 3.00 years). Elastic Net demonstrated the best predictive performance, achieving an MAE of 0.68 ± 0.33 years, RMSE of 1.09 ± 0.80 years, and
Conclusion:
An interpretable regularized linear model accurately reconstructed proprietary device-derived Body Age and outperformed more complex nonlinear algorithms. Body-composition measurements, particularly visceral adiposity and lean tissue, were the strongest predictors of this proprietary device-derived measure of this consumer-oriented measure, whereas graded treadmill performance provided only limited additional predictive value.