Conversion of DXA body composition measurements across different devices

Abstract

BACKGROUND: Dual-energy X-ray absorptiometry (DXA) is widely used to assess body composition, but differences in device calibration, software, and scan modes across manufacturers hinder data harmonization in multi-site studies.

OBJECTIVE: To develop and validate regression models that convert body composition measurements among three DXA systems, GE iDXA, Hologic Horizon A, and Hologic Discovery A, and deploy these models in an accessible web-based tool.

METHODS: A cohort of 101 adults completed same-day whole-body scans on all three machines. Fat mass (FM), lean soft tissue (LST), fat-free mass (FFM), bone mineral content (BMC), and % fat body fat were extracted. Pearson correlations quantified measurement agreement. For each device pair, we developed unadjusted, age-adjusted, and age and sex regression models and evaluated generalizability using leave-one-out cross-validation (Q²). Bland-Altman analyses assessed bias and limits of agreement. A user-friendly web-based application was developed deploy the conversions.

RESULTS: Correlations between DXA machines were high across all body composition measures (% body fat: r = 0.96-0.98; FM: r = 0.98-0.99; LST: r = 0.97-0.99). Adjusted R² values exceeded 0.95 for nearly all regression models, with lower performance observed for FFM conversions involving the Horizon device (adjusted R² = 0.84-0.86). Bland-Altman analysis of % body fat conversions revealed mean bias ranging from 1.02 percentage points for iDXA-to-Discovery to 4.24 percentage points for Discovery-to-Horizon. The Discovery-to-Horizon conversion exhibited a 95% confidence interval entirely above zero [ 1.56,6.91 ] , indicating systematic overestimation of Horizon % body fat. In contrast, the iDXA-to-Discovery conversion showed narrower limits of agreement [ - 0.25,4.52 ] with no evident proportional error. Several conversions demonstrated proportional bias, including Horizon-to-Discovery and Discovery-to-iDXA % body fat predictions, characterized by overestimation at lower adiposity and underestimation at higher adiposity levels.

CONCLUSIONS: Same-day, same-subject measurements across three DXA systems enabled development of machine-to-machine conversion equations with high predictive accuracy for most body composition variables and greater variability for FFM. The accompanying web-based tool provides a practical resource for applying these equations at scale and reducing inter-device measurement differences in multi-site datasets.