Introduction

Bone mineral content (BMC) and non-bone minerals are essential components of body composition. BMC is primarily found in bones and teeth, where it significantly contributes to bone strength. Non-bone minerals, present in body fluids, muscles, and soft tissues, play a crucial role in maintaining overall health. Human BMC peaks in young adulthood, remains relatively stable for a period, and then gradually declines with age1. These age-related changes in bone structure are regulated by multiple interconnected mechanisms, including hormonal imbalances (such as declining levels of estrogen and testosterone), reduced mechanical loading due to decreased physical activity, and the accumulation of senescent cells whose pro-inflammatory secretory phenotype disrupts bone remodeling2. These factors significantly increase fracture risk in the elderly, particularly the incidence of hip fractures. For instance, in men with osteoporosis and specific risk factors, the incidence of hip fractures can be as much as 50 times higher than in men without osteoporosis3. Postmenopausal women, who experience accelerated bone loss, face a lifetime fracture risk of up to 44%4. Accurate monitoring of BMC is particularly important for children and adolescents to assess bone growth and health status. This underscores the significant value of long-term bone health management in preventing bone-related diseases and promoting health across all age groups.

In routine health monitoring and related research, body composition assessment methods are widely applied, including for the detection of obesity5, sarcopenia6, and osteoporosis7. Dual-energy X-ray absorptiometry (DXA) is regarded as the gold standard for clinical bone density assessments, offering high accuracy and repeatability8,9. However, DXA also has several limitations, such as high costs, large equipment size, and the need for professional operators. In contrast, bioelectrical impedance analysis (BIA) estimates body composition by measuring the impedance of weak alternating electrical currents passing through the body. Its advantages include ease of operation, portability, and low cost. Moreover, professional medical personnel are generally not required to perform these measurements. Nonetheless, existing studies have noted that BIA results can be influenced by various factors, including individual physiological differences, racial characteristics, and algorithm variations10,11,12. In addition, BIA cannot directly measure bone mineral content (BMC) but rather indirectly estimates it based on the assumption that the proportions of fat, lean mass, and body fluids remain constant. However, impedance measurement results are affected by individual differences—such as age, race, and hydration status—and equations often fail to fully account for this variability, resulting in significant limitations on the applicability and accuracy of estimation results across different populations.

Despite these limitations, BIA’s portability and noninvasive nature make it a valuable tool in bone health research and community-based screening. In fact, some studies have attempted to use BIA to predict bone mineral content in specific populations and have made certain advancements. For example, Lee et al. evaluated healthy children aged 6–12 years and found that BIA provides reasonable accuracy in estimating total bone mineral content in this age group13. Meanwhile, Lu et al. employed bioelectrical impedance vector analysis (BIVA) to assess both total and regional bone mineral density in older adults, and demonstrated a statistically significant correlation14.

It is understood that BMC prediction equations developed specifically for the Korean population have already been applied in certain BIA devices, such as the ACCUNIQ BC380. However, there is currently a lack of systematic research evaluating the accuracy of these devices in estimating BMC among different age groups within a healthy population. Therefore, this study will recruit healthy participants across a broad range of age groups, measure their BMC using both DXA and BIA, and compare the results to assess BIA’s accuracy and limitations in estimating BMC in diverse populations. Furthermore, this study will preliminarily establish simplified calibration models for different age groups and determine whether optimized BIA prediction models can effectively improve measurement accuracy. Through this study, we aim to provide a convenient and reliable tool for the long-term management of bone health in healthy individuals. Additionally, we will explore the potential of BIA in enabling rapid and accurate bone health screening and enhancing the identification of high-risk populations.

Methods

Subjects

This cross-sectional study was conducted in a kinesiology laboratory from March to September 2023. Male and female participants were recruited through on-campus announcements. Before participation, all subjects provided written informed consent. A strict screening process was implemented to identify eligible participants. Exclusion criteria included electrolyte imbalances, cancer, inflammation, or severe cardiovascular or pulmonary diseases. In total, 148 Korean men (mean age: 24.87 ± 12.43 years) and 154 Korean women (mean age: 34.98 ± 22.24 years) who met the inclusion criteria were enrolled in the study.

Participants were instructed to maintain their usual daily routines and dietary habits, while avoiding alcohol consumption and strenuous exercise for at least 48 h prior to testing. On the day of measurements, subjects were required to fast and refrain from water intake until completion of the assessments. They were also instructed to empty their bladders before testing. Female participants were not measured during menstrual bleeding, and pregnant or potentially pregnant women were excluded from participation. All measurements were performed during daylight hours in a temperature-controlled room (26–28 °C), with participants wearing designated laboratory attire. Height and weight were assessed without footwear using an electronic height and weight measuring instrument (GL-150P, G-Tech International, Gyeonggi-do, Korea). Each participant signed an informed consent form approved by the university’s Institutional Review Board (IRB: 202301-SB-002-01). All procedures conformed to the principles of the Declaration of Helsinki (2008).

Bioelectrical impedance analysis (BIA) measurement

Body composition analysis using BIA was performed with a multi-frequency 8-point tactile electrode system (ACCUNIQ BC380, SELVAS Healthcare Inc., Daejeon, South Korea). Before the measurement, participants used electrolyte wipes to clean their palms and soles. During the test, they stood with their feet apart and arms extended approximately 30° away from the body. Each participant underwent the test three times.

Dual-energy X-ray absorptiometry (DXA) measurement

BMC and soft tissue composition were measured using a fan-beam Discovery Wi (Hologic, Bedford, MA, USA) equipped with Hologic APEX 4.53 software. The scanner used two different energy X-rays to compute the scanned area. Daily calibration of the DXA scanner was performed according to the manufacturer’s guidelines, and experienced radiology technicians positioned the participants on the scanning table.

Statistical analysis

All statistical analyses were conducted using SPSS version 21.0 (SPSS Inc., Chicago, IL, USA). Normality of variables was evaluated using the Shapiro–Wilk test. Bland–Altman plots were generated to assess the agreement between BIA and DXA measurements of fat-free mass (FFM), fat mass (FM), and lean mass (LM), as well as to visually display the bias and limits of agreement.

For normally distributed variables, Pearson’s correlation coefficient was used to assess linear associations between BIA and DXA estimates of non-bone mineral indicators. Differences in participant characteristics across age groups were examined using the Kruskal–Wallis H test. Paired t-tests were used to compare BIA and DXA measurements within the same individuals. Statistical significance was set at p < 0.05.

Using DXA as the reference standard, stepwise regression analysis was initially performed with 32 candidate predictors to develop preliminary predictive models. These candidate predictors included general anthropometric variables (age, height, weight, BMI, FFM, FM, LM, and mineral content), segmental impedance variables (resistance and reactance at low, medium, and high frequencies), and water compartment measures (total body water, extracellular water, and intracellular water) assessed in four body segments: right arm, left arm, right leg, and left leg. Following the stepwise selection, additional manual adjustments of predictor combinations were conducted to explore alternative models and further improve predictive performance. The final optimal model was determined based on the highest adjusted R2 and the lowest root mean square error (RMSE), balancing statistical accuracy and model interpretability.

A power analysis was conducted using G*Power 3.1.9.4 (Universitat Düsseldorf, Germany) to evaluate the adequacy of the sample size (linear multiple regression—fixed model, single regression coefficient). The analysis was based on a medium effect size (f2 = 0.3), α = 0.05, Power = 0.80 and 32 predictors. The minimum required sample size was calculated to be 39 (df = 6), with an actual power of 0.81.

Results

Subject characteristics

The Kruskal–Wallis H test revealed significant differences in subject characteristics among age groups (P < 0.001). Specifically, age, height, weight, Body Mass Index (BMI), Free-Fat Mass (FFM), Fat Mass (FM), and Lean Mass. (LM) showed statistically significant variations across groups (Table 1).

Table 1  Subject characteristics.
Full size table

Reliability assessment of body composition indicators: Bland–Altman and correlation analysis

The Pearson correlation coefficients (r-values) for the indicators (FFM, FM, LM) in all age groups are above 0.96, with the correlation coefficients for FFM and LM approaching 0.99. This indicates a strong linear correlation between BIA and DXA in measuring these body composition parameters (Fig. 1), with all P-values being less than 0.001 (Fig. 1).

Fig. 1
figure 1

Bland–Altman and linear correlation analysis between BIA and DXA for measuring FFM, FM, and LM. Each panel compares BIA and DXA values in four age groups (6–17, 18–35, 36–50, 51–86, and 6–86 years) for LM (left), FFM (middle), and FM (right). Scatter plots of the difference (BIA – DXA) against the mean of BIA and DXA values are presented, with blue and red dashed lines indicating the 95% limits of agreement (± 1.96 SD) and mean difference (bias), respectively. Correlation coefficients (r), P-values.

Full size image

The Bland–Altman plots illustrate the mean differences and the 95% limits of agreement (LOA) between the BIA and DXA measurements across different age groups. The results show that the majority of data points for the LM, FFM, and FM indicators fall within the LOA range, and no significant systematic bias in the difference distribution is observed (Fig. 1).

Optimized regression models for BIA and DXA parameters across different age groups

6–17 Age Group: By incorporating Age, Height, FFM, and Left Leg High-Frequency Impedance​ as the main predictors, the model achieved the highest adjusted R2 (0.90) (Table 2).

Table 2  Predictive performance of optimized models across age groups.
Full size table

18–35 and 36–50 Age Groups: The adjusted R2 values were similar at 0.79 and 0.80, respectively. Mineral and FM were the primary contributing variables (Table 2).

51–86 Age Group: Compared to other groups, the adjusted R2 was lower (0.725). The inclusion of extracellular water (ECW) parameters and total body water (TBW) suggests that water-related variables significantly influence the optimization of models for the elderly (Table 2).

Overall 6–86 Age Group: The adjusted R2 was 0.865, combining Mineral, Left Leg High-Frequency Impedance and Right Leg Medium-Frequency Impedance, and FM, demonstrating a high explanatory power for the model (Table 2).

In the 6–17-year-old group, the optimized model had a mean difference of − 0.02 (95% CI − 0.05287 to 0.01598, P = 0.2868), with a lower limit of − 0.2558 (95% CI − 0.3151 to − 0.1966) and an upper limit of 0.2190 (95% CI 0.1597 to 0.2782). The standard deviation was 0.12, indicating minimal and statistically insignificant paired differences compared with DXA measurements. In contrast, the BC380 equation had a mean difference of 0.46 (95% CI 0.3917 to 0.5354, P < 0.0001), a lower limit of − 0.03222 (95% CI − 0.1559 to 0.09146), an upper limit of 0.9593 (95% CI 0.8356 to 1.0830), and a standard deviation of 0.25, showing significant systematic bias (Fig. 2).

Fig. 2
figure 2

Bland–Altman Comparison of Optimized Model BMC and BC380 BIA Equation BMC with DXA BMC. Each panel presents Bland–Altman plots comparing the BC380 BIA equation-predicted BMC (left) and the optimized regression model-predicted BMC (right) with DXA-measured BMC in four age groups: 6–17 years, 18–35 years, 36–50 years, and 6–86 years. Red and purple dots represent female and male participants, respectively. The mean difference (bias), 95% limits of agreement (± 1.96 SD), correlation coefficient, standard deviation (SD) of differences, root mean square error (RMSE), and statistical significance (P value) are indicated in each panel. The plots show that the optimized regression models consistently demonstrate improved agreement with DXA measurements compared to the standard BC380 BIA equations, particularly in the 6–17 years group.

Full size image

In the 18–35-year-old group, the optimized model had a mean difference of 0.001 (95% CI − 0.03289 to 0.03582, P = 0.9329), a lower limit of − 0.4538 (95% CI − 0.5126 to − 0.3949), an upper limit of 0.4567 (95% CI 0.3979 to 0.5155), and a standard deviation of 0.23, indicating almost no paired difference with DXA measurements. Meanwhile, the BC380 equation had a mean difference of 0.34 (95% CI 0.3034 to 0.3794, P < 0.0001), a lower limit of − 0.1621 (95% CI − 0.2272 to − 0.09704), an upper limit of 0.8450 (95% CI 0.7799 to 0.9100), and a standard deviation of 0.26, showing significant systematic bias and statistical significance (Fig. 2).

In the 36–50-year-old group, the optimized model had a mean difference of − 0.02 (95% CI − 0.07949 to 0.04061, P = 0.5140), a lower limit of − 0.3459 (95% CI − 0.4496 to − 0.2422), an upper limit of 0.3070 (95% CI 0.2033 to 0.4107), and a standard deviation of 0.17, suggesting no significant paired difference with DXA measurements. In contrast, the BC380 equation had a mean difference of 0.25 (95% CI 0.1793 to 0.3199, P < 0.0001), a lower limit of − 0.1325 (95% CI − 0.2539 to − 0.01114), an upper limit of 0.6317 (95% CI 0.5103 to 0.7531), and a standard deviation of 0.19, indicating clear systematic bias (Fig. 2).

In the overall 6–86-year-old group, the optimized model had a mean difference of 0.18 (95% CI 0.1576 to 0.2100, P < 0.0001), a lower limit of − 0.2701 (95% CI − 0.3150 to − 0.2253), an upper limit of 0.6377 (95% CI 0.5928 to 0.6826), and a standard deviation of 0.23, indicating statistically significant differences with DXA measurements. In contrast, the BC380 equation had a mean difference of 0.32 (95% CI 0.2915 to 0.3537, P < 0.0001), a lower limit of − 0.2158 (95% CI − 0.2690 to − 0.1626), an upper limit of 0.8610 (95% CI 0.8078 to 0.9142), and a standard deviation of 0.23, showing clear systematic bias (Fig. 2).

In the 51–86-year-old group, the optimized model had a mean difference of − 0.0003 (95% CI − 0.06908 to 0.06855, P = 0.9939), a lower limit of − 0.4331 (95% CI − 0.5517 to − 0.3146), an upper limit of 0.4326 (95% CI 0.3140 to 0.5511), and a standard deviation of 0.22, suggesting good agreement with DXA measurements and no significant paired difference. In contrast, the BC380 equation had a mean difference of 0.13 (95% CI 0.03404 to 0.2273, P = 0.0093), a lower limit of − 0.4771 (95% CI − 0.6436 to − 0.3107), an upper limit of 0.7385 (95% CI 0.5720 to 0.9050), and a standard deviation of 0.31, showing significant paired differences (Fig. 3).

Fig. 3
figure 3

Comprehensive Error and Agreement Analysis of BMC Prediction Models in the 51–86 Years Group. (A) Bland–Altman plot comparing the BC380 BIA equation-predicted bone mineral content (BMC) with DXA-measured BMC. The red and purple dots represent female and male participants, respectively. The mean difference (bias), standard deviation (SD), root mean square error (RMSE), and correlation coefficient are shown. (B) Bland–Altman plot comparing the optimized regression model-predicted BMC with DXA-measured BMC. (B1) Histogram of standardized residuals from the optimized regression model, showing an approximately normal distribution. (B2) Density plot of residuals versus predicted BMC values for the optimized regression model, with color coding reflecting data density (red: highest density). The local regression (LOESS) trend line shows no significant heteroscedasticity.

Full size image

The residual histogram (Fig. 3.B1) shows that the residuals of the optimized model were normally distributed, with most values concentrated between − 1 and 1, meeting the assumptions of linear regression. The density scatter plot (Fig. 3.B2) further illustrates the relationship between residuals and predicted BMC values. In the optimized model, residuals were mainly clustered around zero and showed a generally symmetric distribution without obvious patterns or trends.

Discussion

This study primarily examined the agreement between Bioelectrical Impedance Analysis (BIA) and Dual-energy X-ray Absorptiometry (DXA) for estimating Bone Mineral Content (BMC) across different age groups. It also evaluated improvements after optimization of the predictive model. The findings showed that BIA’s predictive accuracy for BMC varied among age groups. However, optimization of the model improved predictive performance, especially when tailored to specific populations.

First, we compared the correlation and agreement between BIA and DXA in measuring fat-free mass (FFM), fat mass (FM), and lean mass (LM) across different ages. Both methods showed strong correlations and agreements in these measurements. Despite technical differences in measurement principles and BIA’s sensitivity to population-specific factors15,16, BIA measurements of FFM, FM, and LM generally matched DXA results. This aligns with earlier studies5,17. Previous research suggested that BIA’s BMC estimates depend on its accuracy in measuring FFM18. Therefore, reliable FFM, FM, and LM measurements strengthen the case for BIA’s use in predicting BMC.

Second, BIA and DXA showed good overall correlations for BMC. The optimized regression models included stable general variables, body composition data (FFM, FM, LM, Intracellular water (ICW), Extracellular water (ECW), Total body water (TBW), Mineral), and impedance data. Age-specific equations improved BMC prediction. For instance, in the 6–17-year-old group, adding FFM and impedance at high frequency and medium-frequency raised the adjusted R2 to 0.90 with an RMSE of 0.287, showing better accuracy. Similarly, an earlier study reported an R2 of 0.932 for children aged 6–12 when combining simple measures with FFM and FM13. These results support the accuracy of BIA-based models in younger populations.

In contrast, for adults aged 18–86, the differences between BIA and DXA decreased after optimization. However, predictive accuracy was still lower in adults than in younger individuals. This agrees with findings from earlier work19, which showed that BIA overestimated adult BMC by 3.5–4.7%, with correlations from 0.61 to 0.88. Reduced accuracy in older adults may be due to age-related changes in body composition. Prior studies found increased impedance and resistance after age 6020. Adjusting BIA models for these age-related changes is important. In our study, adding water-related data (ECW and TBW) reduced errors in the 51–86 group. Older people often have lower intracellular water, higher extracellular water, and changes in total body water, reflecting shifts in fluid balance21,22. These changes also make BIA more sensitive to hydration status in older adults23. It is worth noting that although model optimization improved predictive accuracy, errors still remained in the 51–86-year-old group, with the 95% limits of agreement still relatively wide (− 0.43 to 0.43 kg). This suggests that despite enhanced average accuracy, substantial individual differences in BMC predictions persist within this age group. These differences may reflect variations in body composition and hydration status that cannot be fully accounted for by the optimized model. Moreover, differences in gender distribution may also introduce prediction bias. Since men and women differ in fat distribution, muscle mass, and fluid balance, imbalances in the gender sample may further amplify prediction errors, particularly in older age groups. These factors point to an important limitation of the current BIA model.

In the overall model (6–86 years), optimization raised the adjusted R2 to 0.87 and lowered the RMSE to 0.295. Still, large individual differences between BIA and DXA results remained. Predictive equations tailored to different demographic groups are necessary to reduce these individual differences and improve accuracy.

Variations in hydration status in growing children and adolescents challenge BIA models. Rapid changes in body water, proteins, minerals, and limb-to-trunk proportions affect BIA assumptions during growth24,25. Using adult equations for children or adolescents can mislead results. Age-specific calibration improves accuracy26. Adults have relatively stable total body water27, and a previous study has confirmed the accuracy of BIA in adults28. However, older adults have lower muscle mass, higher body fat, lower total water, and higher ECW/ICW ratio29,30,31,32. These factors challenge BIA model assumptions. Studies show that failing to include age effects reduces accuracy23,33. A recent review also note that missing population-specific models affects precision34. BIA is also sensitive to hydration. Water retention or loss can shift BIA results35. Overhydration (e.g., drinking a lot in a short time) biases BIA data36. Dehydration also causes errors37.

Despite these limitations, BIA remains a cost-effective and simple tool for estimating BMC in settings with fewer resources. While tailored models improve predictions, clinical use of BIA still has limits. Current BIA models estimate whole-body or limb BMC. They cannot replace DXA for full or site-specific bone mass diagnosis (e.g., lumbar spine, femoral neck)38. Thus, BIA should be used carefully in clinical settings. It is best for indirect measurements, screening, or tracking bone mass changes. High-risk individuals identified by BIA should have DXA scans for confirmation. Standardized procedures are also needed to limit errors from hydration changes. Further external validation across devices and groups is also essential. BIA is most useful in community and primary care settings. For those with noticeable bone changes, DXA should confirm the results. This stepwise approach helps meet clinical management needs.

Study limits

This study has several limitations. First, only a single BIA device model was used. Its specific frequency parameters and built-in algorithms may not represent the performance of all BIA devices. Differences in impedance measurements and calibration equations across manufacturers and models may affect the comparability and generalizability of the results. Second, the study did not include external validation with chemical analysis, the gold standard for measuring BMC. Although DXA is considered a reliable reference, the lack of chemical analysis limits further confirmation of absolute BMC values. Third, although the total sample size was 302, the age distribution was not balanced, which may limit the applicability of the model across different age groups. The 36–50-year-old group had too few participants to be representative, while the 51–86-year-old group encompassed a wide age range and substantial inter-individual variability, potentially affecting the model’s robustness among older adults. The 6–17-year-old group had a relatively balanced sample size; however, due to rapid growth and dramatic changes in body composition during adolescence, BIA predictions in this group still require larger samples and longitudinal validation. Furthermore, the study population predominantly comprised healthy and relatively young individuals, leading to a high degree of sample homogeneity that may further restrict the external validity of the findings. Models developed from homogeneous samples may not adequately reflect the diversity present in broader clinical or community populations, thereby limiting the applicability of the BIA equations.

In addition, the gender distribution was not fully balanced. This could lead to differences in model performance between men and women. Men and women have significant physiological differences in fat distribution and hydration status. These differences may be magnified in smaller sample groups of middle-aged and older adults, which could increase prediction bias in BIA equations. Finally, this study was conducted exclusively in a Korean population, which may limit the generalizability and applicability of the findings to other ethnicities or populations. Therefore, future research should address these limitations by including multiple BIA devices, more diverse samples, and cross-validation with gold-standard chemical analyses to improve external validity and scientific rigor.

Conclusion

In conclusion, the findings demonstrated that BIA performed well in estimating key body composition parameters when compared to DXA. By incorporating body composition indicators to optimize prediction models for specific groups, the accuracy of BMC predictions improved significantly. These results highlight the potential of BIA as a practical, portable, and cost-effective alternative for community-based bone health monitoring. However, the applicability to specific populations should be prioritized during implementation to minimize errors and ensure reliability. Nevertheless, DXA remains the indispensable gold standard for clinical bone health assessment, particularly for high-risk populations.