New GNR Method Estimates Body Composition Non-Invasively.

Accurate assessment of body composition, including metrics like body fat percentage (BFP), bone mineral density (BMD), and appendicular lean mass (ALM), is vital for evaluating overall metabolic, skeletal, and muscular health. Traditional direct methods, such as dual-energy X-ray absorptiometry (DXA), involve specialized equipment and radiation exposure, limiting their accessibility and frequency of use. To overcome these limitations, a new framework called $p$SADE-GNR (target-aware, state-adaptive $p$-Dirichlet graph neural regression) has been developed. This method estimates body composition outcomes using only non-invasive anthropometric measurements. It employs a neural encoder to map participant data to hidden states, which are then propagated over an outcome-specific participant-similarity graph using a state-adaptive forward-Euler discretization of the graph $p$-Dirichlet energy flow. The model's effectiveness was demonstrated using clinical data, where a correlation-weighted version, utilizing original standardized measurements, achieved the lowest root mean squared error across nine primary outcome-cohort combinations. It also outperformed established support vector regression or least-squares support vector regression benchmarks in eight of these nine comparisons. This research positions $p$SADE-GNR as a highly promising non-invasive alternative for body-composition estimation.