The present study measured FFM in 544 healthy Asian individuals using DXA (297 male, 257 female; age range, 16–75 years). We used BIA to measure the impedance at 50 kHz of the lower extremities in a standing position to develop a multivariable model for predicting FFM using DXA measurements. We also evaluated the validity of the developed multivariable model with a double cross-validation technique and assessed the accuracy of the model in an all-subjects sample and in different BF% subgroup samples. The results of the study indicated that the FFM predictive model based on BIA estimates is a valid method for assessing FFM in healthy subjects with different BF% values. The force of gravity has an effect on the fluid distribution in our body. Depending on the body position, gravity may also cause differences in blood pressures. As a result, regulation of blood volume may become challenging: standing still leads to rapid and persistent plasma volume loss of up to 7 % for a 30-min period . Nunez et al.  performed foot-to-foot standing upright and supine position impedance measurements and obtained the following results. There was a high correlation between upright and supine position impedance measurements of the lower extremities. The difference in impedance measured by the two methods versus the mean impedance for the two methods was evaluated as a Bland–Altman plot (r = 0.44, p = 0.23, NS). The plot showed a small but systematic difference between the two methods. In a study by Rush et al. , the foot-to-hand impedance decreased by up to 9 ohm (mean, 5 ohm; 1.0 %) over 10 min of standing and increased by up to 7 ohm (mean, 3 ohm; 0.7 %) in the lying position. Based on the results of both studies, the difference in the impedance measures was caused by the changes in the effects of gravity on the different positions and body fluids. Oshima  reported that the average foot-to-foot impedance value would decrease by 6.8 % after 6 h of continuous measurements. Kushner et al.  also reported a −3 % to 1 % change during a 5-min standing upright position and 10-min supine position in hand-to-foot impedance measurements. In the present study, regardless of standing or lying down, only a 1 % decrease in the foot-to-foot impedance measurement occurred over a 3-min period (data not shown).
The standing foot-to-foot BIA method described herein produced inconclusive results. The present study had the following characteristics: (a) We used the same instruments in the same setting to measure the impedance by foot-to-foot BIA in the standing position and FFM by DXA in a large, single-institution Asian sample; such patients have been insufficiently studied in BIA research to date. (b) Instead of evaluating the validity of existing commercial instruments, this study aimed to develop an FFM predictive model using standing foot-to-foot BIA. (c) This study tested not only the accuracy and suitability of BIA for assessing body composition in a general population, but also its performance in subjects with different BF% values.
The present study used h2/ZF-F and other anthropometric variables, such as weight, sex, and age, as predictive variables to develop the prediction model. We used h2/ZF-F instead of h2/reactance (XC) or resistance (R), as adopted by other studies [23, 30], as a predictor in the regression model for the following reasons: The QuadScan 4000 produces a 50-kHz frequency and provides measured results for resistance, impedance, and reactance. Nunez et al.  proposed a standing foot-to-foot bioimpedance analysis FFM estimate model in which they used Z (impedance) as the estimate variable; this variable has also been widely used in other studies of BIA standing-position body composition estimation . Furthermore, in similar research, Kotler  and Lukaski et al.  indicated that when using resistance or impedance estimating FFM, TBW (total body water) and TBK (total body potassium) have no significant difference. For these reasons, we used impedance as the estimated variable. When developing a body composition predictive model, the predictors should be easy to measure, accurate, reproducible, and physiologically related to the dependent variable [34, 35]; the predictors in our model meet these requirements [36, 37].
When developing a regression model, the following issues should be taken into consideration to avoid violating assumptions and to ensure that the regression analyses have sufficient power: collinearity, sample size, number of predictors, cross-validation, and SEE . When additional variables were added into the FFM predictive equation using BIA and other anthropometric variables, collinearity was present if the predictive variables were highly correlated. This might affect the estimation of the regression coefficient in a predictive model, which may lead to incorrect identification of the predictive variables. The VIF analysis was conducted to identify potential problems related to collinearity. When the VIF of the predictors had exceeded ten, the collinearity was considered to be severe. In this study, the VIF was smaller than five; thus, collinearity did not exist.
To ensure sufficient power, the minimum sample size was set to 91 when the effect size was medium, the number of predictors was five, the power was 0.8, and the alpha value was 0.05 . When the effect size was small, the minimum sample size was 686. Although the number of subjects in the present study was less than 686, the sample was large enough to minimize variable inflation and improve reliability. Additionally, a cross-validated technique was used to validate the prediction model. In this study, height2/impedance, sex, weight, and age were the predictors in the regression model. In the model, the correlation coefficient of height2/impedance and FFMDXA was 0.92, and the standardized coefficient β was 0.43; these values explain approximately 43 % of the variance of FFM. The prediction model developed using the G1 and G2 data was double cross-validated with an RMSE of 2.31 and 2.18 kg, respectively and a bias ± SD of −0.01 ± 3.22 and 0.05 ± 3.12 kg, respectively. The similar results derived by the G1 and G2 models indicate that the predictive models can accurately predict FFM. The all-subjects predictive model also showed results similar to the G1 and G2 models in terms of the correlation coefficient, SEE, and CV, thus validating the accuracy of the predictive model. We also randomly assigned subjects into two groups to double cross-validate the regression models. The results were similar to the BF%-matched samples; the regression lines of FFMBIA against FFMDXA developed using randomly assigned data sets demonstrated a similar trend that deviated from an identical line (slope = 0.93 for G1 and 0.92 for G2; data not shown).
When comparing the results of our predictive model with those of previously published studies on supine-position hand-to-foot BIA measurements, the correlation coefficient and SEE were similar to those of Kotler et al.  (r2 = 0.83, n = 256, SEE = approximately 3.0 kg), Sun et al.  (r2 = 0.92, n = 1095, RMSE = 2.9 kg), Heitmann et al.  (r2 = 0.90, n =139, SEE = 3.6 kg), and Sun et al.  (r2 = 90, n = 734, RMSE = 3.9 kg); however, they were lower than those of Kyle et al.  (r2 = 0.96, n = 343, SEE = 1.8 kg) and Deurenberg et al.  (r2 = 0.92, n = 661, SEE = 2.6 kg). These results may have been because the correlation coefficient of the predictive value and the measured FFM value tended to be higher in the hand-to-foot model than in the foot-to-foot model [42, 43]; this may be a shortcoming of the foot-to-foot model in assessing FFM.
The Geneva BIA equation published by Kyle et al.  provides ideal results of a high r2 and low SEE (r2 = 0.96, LOA = −3.4 to 3.5 kg, and SEE = 1.72 kg); however, their subjects’ BMI range was narrower (17.0–33.8 kg/m2) than that of the subjects in our study (15.9–43.1 kg/m2). Sun et al.  and Deurenberg  indicated that the estimated results were affected by the level of adiposity. The developed predictive equations in our study overestimated FFM in subgroups with a higher BF% (male, BF%DXA > 30 %; female, BF%DXA > 40 %). When these subjects were excluded and used to develop another model, then the results (n = 442; BMI, 15.8–36.9 kg/m2; r2 = 0.94; SEE = 2.80 kg, LOA = not reported) were comparable with those reported by Kyle et al. . Although these results may be appealing, they have limited application. We included subjects with a high BF% to broaden the application range. The average FFMDXA of the subjects in our study was 50.17 ± 11.25 kg, while that in the Geneva BIA equation was 54.0 ± 10.5 kg. Their sample had a smaller SD for FFMDXA, indicating that their data tended to be closer to the mean, resulting in a smaller SEE. The standing foot-to-foot impedance measurement may be convenient, but has a significantly smaller FFM correlation than the hand-to-foot impedance measurement [31, 42]. Based on the estimate equation suggested in our study, the LOA may be large; however, we consider it acceptable. This is one of the limitations of the present study. When estimating FFM using our predictive equation in subjects with a high level of adiposity (female, BF%DXA > 40 %; male, BF%DXA > 30 %), the bias ± SD in female and male subjects was 2.0 ± 2.9 and 2.1 ± 3.2 kg, respectively. Although the bias and SD were higher than those in the other leaner subgroups in our study, the results show that our predictive equations performed better for estimating FFM in subjects with a high BF% than did the equation developed by Jakicic et al.  (r2 = 0.66, SEE = 8.8 kg, n = 123, and LOA = not reported).
Age has not been included as a predictor in every model in other published studies [23, 36, 41]. Some models excluded age because it only explained limited variance in FFM . However, in our predictive model, age explained approximately 18.0 % of the variance in FFM and was therefore included in the predictive model. Several studies have indicated that the concentration of potassium in fat-free tissue decreases systematically with age [46, 47]. There are important age-related changes in the composition of FFM. The main molecular components of the FFM are water, protein, osseous and nonosseous mineral, and glycogen. The proportion of water, protein, and osseous mineral in the FFM vary systematically with age. Kyle et al.  examined the accuracy of a predictive model with different age groups. Many studies have reported that the accuracy of BIA estimation is affected by the level of obesity [18–20]. Therefore, this study examined the accuracy of a model for predicting FFM in individuals with different percentages of body fat. The predictive value of FFM using our model was not significantly different from FFMDXA among the subgroups of different BF% values and sexes, and the correlation coefficients were 0.87 (p = 0.92) in females and 0.89 (p = 0.97) in males. These results indicate that BIA is an accurate method for assessing FFM in individuals with a BF% in the range evaluated in our study. Moreover, standing foot-to-foot BIA can be used as a convenient method to assess the different BF% values in male and female adults. Clinical use of BIA in patients with abnormal hydration cannot be recommended until further validation has proven that a BIA algorithm is accurate in such conditions. The present study focused on the different foot-to-foot BIA BF% values; we did not discuss differences in foot-to-foot BIA FFM estimate measurements based on either regional composition or different body types; these are topics requiring further discussion. Moreover, in patients with body shape abnormalities, very small or large body heights, or relative sitting heights, the use of prediction equations in subjects with an abnormal body build (e.g., acromegaly or amputation) should be interpreted with caution . Many published studies on BIA estimate equations have used impedance as an estimate variable, but the present study applied the impedance variable in the standing foot-to-foot model and found satisfactory results for estimating FFM in a healthy Taiwanese population (BMI = 16–43 kg/m2).