Predicting DXA-Derived Body Composition in Older Women Using Anthropometry-Based Linear Regularized Regression

This study aimed to develop and validate Elastic Net-based predictive models using simple anthropometric measurements to estimate dual-energy X-ray absorptiometry (DXA)-derived body composition variables in older women.

Body mass, skinfolds, circumferences, and DXA measures of total and regional fat mass (FM) and lean soft tissue (LST) were obtained from 523 older women. Elastic Net regression developed models for total body, upper- and lower-limb FM and LST, and android and gynoid FM. The sample was split into training, validation, and independent testing sets.

Elastic Net models demonstrated high predictive accuracy for DXA-derived body composition. Coefficients of determination were strong for total FM and LST (R² ≥ 0.884), and good for regional compartments (R²≥ 0.711). Root mean square errors were low across outcomes (0.272-1.858 kg). Bland-Altman analyses showed minimal mean bias for most variables, except for lower-limb FM and android FM, which presented small but significant biases (≤ 2.78 %). Proportional bias was identified in some models, although the explained variance remained low (≤22%). Limits of agreement were narrow for total body estimates (≤11.73%), while wider limits were observed for segmental measures.

Elastic Net-based models provide accurate estimates of total body composition and acceptable estimates of regional compartments, although regional predictions showed greater individual variability than total body estimates. However, these prediction equations should be interpreted as internally validated models intended primarily for screening and research applications.