Validity and reliability of a mobile digital imaging analysis trained by a four-compartment model

Digital imaging analysis (DIA) estimates collected from mobile applications comprise a novel technique that can collect body composition estimates remotely without the inherent restrictions of other research-grade devices. However, the accuracy of the artificial intelligence used in DIA is reliant on the accuracy of the developmental methods. Few DIA applications are trained by multicompartment models, but this developmental strategy may be most accurate. Thus, the aim of the present study was to assess the precision and agreement of a DIA application with developmental software trained by a four-compartment (4C) model using an actual 4C model as the criterion method.

For this cross-sectional study, body composition estimations were collected from 102 participants (63 females, 39 males) using the methods necessary for a rapid 4C model and a DIA application using two different smartphones.

Intraclass correlation coefficients (0.96-0.99; all p < 0.001) and root mean square coefficients of variation (0.5%-3.0%) showed good reliability for body fat percentage, fat mass and fat-free mass. There were no significant mean differences between the 4C model or the DIA estimates for the total sample, by sex, and for non-Hispanic White (n = 61) and Black/African-American (n = 32) participants (all p > 0.050). DIA estimates demonstrated equivalence with the 4C model for all variables but revealed proportional biases that underestimated body fat percentage (both β = -0.25; p < 0.001) and fat mass (both β = -0.07; p < 0.010) at higher degrees of each variable.

DIA applications trained by a 4C model are reliable and produce body composition estimates equivalent to an actual 4C model.