Total and regional appendicular skeletal muscle mass prediction from dual-energy X-ray absorptiometry body composition models

Sarcopenia, sarcopenic obesity, frailty, and cachexia have in common skeletal muscle (SM) as a main component of their pathophysiology. The reference method for SM mass measurement is whole-body magnetic resonance imaging (MRI), although dual-energy X-ray absorptiometry (DXA) appendicular lean mass (ALM) serves as an affordable and practical SM surrogate. Empirical equations, developed on relatively small and diverse samples, are now used to predict total body SM from ALM and other covariates; prediction models for extremity SM mass are lacking. The aim of the current study was to develop and validate total body, arm, and leg SM mass prediction equations based on a large sample (N = 475) of adults evaluated with whole-body MRI and DXA for SM and ALM, respectively. Initial models were fit using ordinary least squares stepwise selection procedures; covariates beyond extremity lean mass made only small contributions to the final models that were developed using Deming regression. All three developed final models (total, arm, and leg) had high R2s (0.88–0.93; all p < 0.001) and small root-mean square errors (1.74, 0.41, and 0.95 kg) with no bias in the validation sample (N = 95). The new total body SM prediction model (SM = 1.12 × ALM – 0.63) showed good performance, with some bias, against previously reported DXA-ALM prediction models. These new total body and extremity SM prediction models, developed and validated in a large sample, afford an important and practical opportunity to evaluate SM mass in research and clinical settings.

most applicable and safe reference method when applied in healthy adults 6 .Moreover, MRI additionally provides regional estimates of SM mass and structure 7 .However, acquiring a whole-body MRI scan is costly and automated image analysis methods are not widely available outside of proprietary vendors.Clinical studies of SM thus tend to have relatively small sample sizes, thereby limiting statistical power and generalization of results.
A widely embraced alternative to whole-body MRI for measuring muscle mass is dual-energy X-ray absorptiometry (DXA) 8,9 .A large proportion of SM is distributed in the appendages, a region in which DXA can quantify the amount of lean soft tissue present.Kim et al. in 2002 exploited this anatomic relationship and imaging capability by reporting SM mass prediction equations with DXA-measured appendicular lean soft tissue mass, now referred to as appendicular lean mass (ALM), as a key predictor variable 9 .Whole body MRI served as the reference method for measuring total SM mass in Kim's study that included a racially/ethnically mixed model development sample of 321 adults.Kim et al. reported their findings again 2 years later 8 following a reanalysis of 270 MRI scans that included removal of intermuscular adipose tissue (IMAT) from the SM estimates.The studies reported by Kim et al. 8,9 did not report SM mass prediction models for regions such as the arms and legs.Studies that have followed Kim et al. identified additional SM predictor variables such as regional and total body fat mass 10,11 on small (< 70) adult samples.Adipose tissue has a small amount of lean mass that contributes to the total ALM as measured by DXA that can lead to an overestimate of total SM mass in people with obesity 12 .Selected groups, such as young athletes, may also have ALM-SM relations that differ from those of non-athletic older adults 10 .
A series of studies over the past two decades at the Institute of Human Nutrition, Kiel University, Germany included detailed MRI total-body and regional measurements of SM and other organ and tissue volumes.Participants additionally had total-body DXA scans.The large available Kiel sample of 475 participants provides the important opportunity to develop new total body and extremity SM mass prediction equations and secondarily to validate Kim's original total body SM prediction model 8 .

Results
Baseline sample characteristics.The full evaluated sample of 216 men and 259 women had a mean age of about 50 years and a BMI of 26 kg/m 2 (Table 1).The BMI and age distributions across the model development and validation samples were similar on women and men and included wide ranges of both BMI and age.These characteristics, including MRI estimates of total body SM, are similar to those reported by Kim et al. 8 (Supplementary Table 1).Table 1.Participant Characteristics (X ± SD).ALM appendicular lean mass, BMI body mass index, DXA dualenergy X-ray absorptiometry, MRI magnetic resonance imaging, N number, SM skeletal muscle, UW, NW, OW, OB are underweight (BMI, < 18.5 kg/m 2 ), normal weight (18.5-24.9kg/m 2 ), overweight (25.0-29.9kg/m 2 ), and obese (> 30 kg/m 2 ).The distribution of MRI-measured SM and DXA-measured lean mass across the total body and regions for participants in the current study is shown in Table 1.As expected, men had more total and regional SM mass than women (~ 30 vs. 20 kg).Of total SM mass, 13.0%, 36.5%,50.5%, and 63.5% was present in the arms, trunk, legs, and appendages, respectively, in the women.Corresponding results in the men were 14.6%, 37.0%, 48.4%, and 63.0% of total body SM was present in the arms, trunk, legs, and appendages, respectively.
The proportions of DXA-measured lean mass as MRI-measured SM were largest in the legs (~ 0.70) and smallest in the trunk (~ 0.30) (Table 2); the proportion of ALM as SM was about 0.70.Men had a larger proportion of lean mass as SM in their trunk and legs and less in their arms compared to the women, although none of the regional differences were statistically significant.The proportion of total lean mass as SM was about 0.46 with the level larger in men (0.48) than in women (0.45; p = NS).
Model development and validation.The contributions of covariates on total, arm, and leg SM estimation beyond extremity lean mass components were negligible and hence the following analyses were conducted using Deming regression.The full least-squares analyses are presented in Supplementary Table 2.

Total body.
The strong association between MRI-measured total body SM mass and DXA-measured ALM is shown for the full sample in Fig. 1 (R 2 , 0.93; p < 0.001).The total body SM mass prediction model (Table 3, Fig. 2A) had a validation R 2 of 0.93 and RMSE of 1.74 kg.No significant difference from the line of identity was observed for the slope (95% CI: 0.91, 1.02) and intercept (95% CI: −0.56, 2.35).Additionally, statistical equivalence was demonstrated (p < 0.001) between MRI-SM and predicted SM mass using regions of 2.5% of MRI-SM (0.64 kg).No significant proportional bias was observed in the Bland-Altman analysis (slope 95% CI: −0.06, 0.06; Fig. 2B).Sex-specific models and their performance are presented in the Supplementary Tables 3 and 4 and Supplementary Figs. 1 and 2).Table 2. Proportions of DXA-measured lean mass as MRI-measured SM (kg/kg; X ± SD).None of the mean differences between women and men were statistically significant.DXA dual-energy X-ray absorptiometry, MRI magnetic resonance imaging, SM skeletal muscle.Vol:.( 1234567890) Arm.The arm SM mass prediction model (Table 3, Fig. 2E) had a validation R 2 of 0.88 and RMSE of 0.41 kg.
SM predicted by Kiel equation vs. SM measured by Kim et al. 8 .ALM measured by Kim et al. 8 was used to derive a total body SM estimate using the newly developed Kiel SM prediction model.The Kiel-predicted SM values were then compared to MRI-measured SM by Kim et al. 8 Cross-validation of the Kiel equation indicated a significant difference from the line of identity for the slope (95% CI: 0.88, 0.92) and intercept (95% CI: 1.24, 2.32) (Fig. 3C); the RMSE was 1.63 kg with an R 2 of 0.96.Statistical equivalence was not observed (p = 0.44) using regions of 2.5% of MRI SM (0.60 kg).Significant proportional bias was observed (95% CI for slope: −0.11, −0.06) in the Bland-Altman plot (Fig. 3D).

Discussion
New total body and extremity SM mass prediction equations were developed and validated in the current study using MRI and DXA data from a large sample evaluated at Kiel University's Institute of Human Nutrition.Additionally, we examined the widely used Kim SM prediction model in the current sample and cross-validated the newly developed total body SM prediction equation with Kim's original MRI and DXA data.
Several observations emerge from the current study results and the other previously reported smaller scale and more limited studies 10,11 aimed at developing total body SM prediction equations that are summarized in Table 4. First, we again confirmed that about 60-70% of total body SM is present in the extremities and that extremity lean mass (i.e., ALM) is about 70% SM.It's not surprising, therefore, that total body SM and ALM are highly correlated with each other; the univariate regression R 2 in the current study for SM versus ALM was 0.93 (p < 0.001) and in Kim's study 8 0.96 (p < 0.001).Similar strong associations between SM and ALM were observed by Zhao et al. in 66 Chinese men and women (R 2 , 0.97, p < 0.001) 11 and to a less extent by Sagayama in 30 young athletic Japanese men (r, 0.885, p < 0.001; ref 10) 10 (Table 4).By contrast to ALM, trunk lean mass was only onethird SM, the rest presumably visceral lean tissues such as liver, kidneys, and heart.These observations thus again affirm that DXA ALM is an excellent starting point for developing total body SM prediction equations.
Previous studies have included additional covariates beyond ALM in SM prediction models.Specifically, for the total body SM prediction models, age, sex, muscle distribution, and %fat appear as covariates in earlier prediction equations shown in Table 4.These observations may be expected as the developed SM prediction models are empirical and capture associations that likely reflect small variations in SM distribution as a function of sex, age, athletic "fitness", and variation in non-muscle lean mass composition 13 .However, in the present analysis, the contribution of such covariates on SM estimation was negligible (Supplementary Table 2).As such, a simplified model using ALM as the sole predictor of SM and accounting for errors in both MRI and DXA estimates was developed.Similarly, the developed leg and arm SM estimation equations use solely leg LM and arm LM as  8 Collectively, ALM and extremity lean are the main predictors of total and regional SM, respectively, which  www.nature.com/scientificreports/allows for the development and utilization of parsimonious prediction models.However, small contributions to between-individual variance can be made by other covariates in some instances.
In the current study we also observed that SM predicted by Kim's model and the models reported by Zhao et al. 11 and Sagayama et al. 10 did not predict identical values to those estimated from our new total body SM model.We found small absolute differences and some bias in predicted SM between our model and Kim's model, 8 even when using Kim's original data.Moreover, we similarly found strong correlations but absolute differences and bias when applying Zhao and Sagayama's models 10,11 to the Kiel dataset (data not shown).These kinds of SM prediction variation can be anticipated due to between-sample, DXA system 14 , MRI scanner, and image segmentation method differences (Table 4).To explore the magnitude of potential DXA scanner differences, we compared Hologic Discovery and GE Lunar iDXA estimates of ALM in our laboratory (Supplementary Fig. 3), similar to the study reported by Park et al. 14 that compared Hologic Horizon and GE Lunar Prodigy scanners.Although ALM measured by both DXA systems in our laboratory were highly correlated (n = 45; R 2 , 0.99), average scanner differences (X ± SD; 0.54 ± 0.58 kg; p < 0.001) and significant bias (p < 0.05) were present.Park et al. 14 found between-scanner ALM differences of 1.79 ± 0.92 kg (p < 0.001).These kinds of between DXA and MRI system measurement differences are likely part of the reason why we found small mean differences and bias between our SM predictions and those of Kim et al. 8 A concern raised in several previous publications is that DXA ALM, and thus predicted SM, is not "true" muscle mass 15 .As noted earlier, ALM is linked to SM through several additional covariates.Thus, for example, when people with obesity lose weight some of the changes observed in ALM may be accounted for by changes in the lean portion of appendicular adipose tissue.Preferential changes in SM distribution with interventions might also impact predictions with empirical total body SM equations.Lastly, DXA-predicted SM and SM measured with MRI share in common an evaluation of "total" wet muscle that includes tendons, nerves, blood vessels, and connective tissues.Other methods, such as D 3 -creatine dilution 15 , multifrequency bioimpedance analysis 16 , and ultrasound 17 can be used to derive estimates of muscle "quality" that go beyond an evaluation of the total intact muscle.
Although the current study models were developed on the largest sample to date, ideally much larger and more diverse samples should be evaluated in the future.This limitation will likely be overcome when automated MRI analysis software becomes available, thus reducing study analysis cost and execution time.Our study participants were all Caucasian, and thus generalizing our SM prediction models to other race and ethnic groups optimally should include a priori validation.Our models also did not include potential covariates such as fitness level or type and duration of exercise training.Participants in the current study were also healthy and fully functional and thus a need exists to expand model development samples to people at the extremes of muscularity such as patients with conditions such as sarcopenia on the one hand and body builders on the other.Lastly, we limited our participants to those with BMIs < 35 kg/m 2 .While it is feasible to conduct DXA and MRI scans in people with higher BMIs, measurement errors increase at and above this level of adiposity and thus we elected, as did Kim et al. 8 , to stay with BMIs < 35 kg/m 2 .
The current study advanced a new set of total body and extremity SM prediction equations that should be useful in the study of conditions related to variation in muscularity.Even the relatively small arm lean component yielded a good arm SM prediction equation with a small RMSE of 0.41 kg.Our models are founded on a large sample of healthy adults ranging in age and BMI.The current study findings also suggest the need to standardize DXA and MRI measurement methods and analyses across centers with the aim of creating universal DXA SM prediction models.Combining predicted values for total SM with other estimates of muscle "quality" provides an important opportunity for future research.

Figure 1 .
Figure 1.Total skeletal muscle mass (SM) measured with MRI versus appendicular lean mass (ALM) measured with DXA in the whole sample (n = 475).The Deming regression equation, line (solid), and R 2 are shown in the figure (p < 0.001).

FFigure 2 .Figure
Figure 2. Predicted total, arm, and leg skeletal muscle (SM) mass versus corresponding value measured with MRI in the validation sample (n = 47 women; 48 men) on the left (A,C,E) and associated Bland-Altman plots on the right (B,D,F).The regression equations, lines, R 2 s, and 95% limits of agreement (LOA) are shown in the figures.The statistical significance of each panel is summarized in the text. https://doi.org/10.1038/s41598-023-29827-y The leg SM mass prediction model (Table3, Fig.2C) had a validation R 2 of 0.91 and RMSE of 0.95 kg.No significant difference from the line of identity was observed for the slope (95% CI: 0.90, 1.02) or intercept (95% CI: −0.35, 1.30).Additionally, statistical equivalence was demonstrated (p < 0.001) between MRI leg SM and predicted leg SM using regions of 5% of MRI arm SM (0.61 kg).No significant proportional bias was observed in the Bland-Altman analysis (slope 95% CI: −0.06, 0.07; Fig.2D).

Table 3 .
Developed SM mass prediction models.ALM and lean mass units are in kg.ALM appendicular lean mass, RMSE root mean square error, SM skeletal muscle.

Table 4 .
Comparison of previously reported and current SM prediction models.Mass units, kg; Sex, 0, woman; 1, man.A age, ALM appendicular lean mass, BMI body mass index, M men, SM skeletal muscle, W women.