Comparison of adult height prediction using bone age and body composition for growth assessment in Korean children

Abstract

To compare adult height (AH) predictions using body composition-based biological age with those derived from bone age in Korean children. A multicenter, assessor-blinded, prospective study was conducted with 80 healthy children aged 7–13 years. Participants were assessed using two methods: the traditional Tanner-Whitehouse 3 (TW3) bone age method and a model based on artificial intelligence (AI), incorporating body composition metrics such as BMI, fat-free mass, and muscle mass through bioelectrical impedance analysis. The clinical equivalence between the two prediction methods was evaluated, with a non-inferiority margin of 0.661 years. The difference in predicted bone age between the AI-based method and the TW3 method was 0.04 ± 1.02 years, indicating clinical equivalence. Exploratory analysis showed a positive correlation between lean mass and bone age, suggesting that body composition metrics could reflect skeletal maturity. Therefore, the AI-based method utilizing body composition parameters was clinically equivalent to the traditional TW3 method for predicting AH. This approach offers a viable alternative for predicting adult height in pediatric populations, emphasizing the potential for integrating personalized metrics such as body composition into routine growth monitoring; however, further research is needed before it can be widely applied in clinical practice. Future studies should explore its utility in children with growth disorders and refine the model across different growth phases.

Introduction

Predicting adult height (AH) in a growing child or adolescent is crucial not only in routine clinical practice but also in clinical research, such as estimating height gains from growth-promoting treatments. There are several methods for predicting adult height, with the most commonly used being the Bayley-Pinneau (BP) method¹, Roche-Weiner-Thissen (RWT) method², and Tanner-Whitehouse (TW) method³. Among these, the Tanner-Whitehouse 3 (TW3) method is known to have a relatively small prediction error because it replaces bone age with a skeletal maturity score, which is less affected by secular trends. However, the TW3 method also has limitations that impact its application in clinical settings. These include significant interobserver and intraobserver variability, which can lead to inconsistencies in bone age assessments⁴. This variability is especially problematic during late puberty when hand bone growth slows, making precise assessment challenging⁵. Additionally, the TW3 method requires radiographic imaging and specialized analysis, which may increase resource use and limit its accessibility for frequent monitoring³.

Recent studies suggest that body composition metrics, such as fat-free mass (FFM), BMI, and body fat percentage, are essential for understanding growth patterns beyond skeletal measurements^6–11. These metrics, influenced by nutrition and environment, offer an alternative means of predicting AH. According to reports exploring the relationship between body composition and growth, graphs created by selecting reference values and percentiles for each body composition index in normal adolescent populations can be helpful in assessing growth and health risks¹². In particular, a study that extensively examined FFM by height in subjects ranging from 5 to 59 years found that height was the most important factor contributing to FFM, which could be represented by a nonlinear regression model¹³.

Despite the growing understanding of the relationship between body composition and growth, no studies have reported the use of body composition for predicting adult height in Korean children. Therefore, in this prospective study, we aimed to compare adult height predictions from an AI-based model incorporating BMI and body composition metrics with those from the TW3 method in normal children. We hypothesized that there would be no significant difference between the two methods in predicting AH.

Methods

Study design

This study was conducted in accordance with the Declaration of Helsinki, and this was a multicenter, assessor-blinded, prospective, controlled, device comparison confirmatory trial conducted to evaluate the clinical equivalence between two methods of bone age assessment: the GP Solution (software-based assessment)^6,11 and the traditional X-ray-based TW3 method³. The trial aimed to determine whether the AI-based software, GP Bio Solution, provides results comparable to the TW3 method, commonly used in clinical practice. The study was conducted at two institutions: Kyung Hee University Hospital and Korea University Hospital (IRB No: 2023AN0417).

Inclusion and exclusion criteria

The inclusion criteria were children aged 7–13 years who were undergoing their first bone age assessment. Additionally, written informed consent was obtained from both the children and their legal guardians. Exclusion criteria included children who were receiving growth hormone treatment, children diagnosed with precocious puberty, or any children with chronic diseases or genetic conditions that could affect growth, such as Down syndrome, Turner syndrome, or metabolic disorders.

Participants

A total of 80 healthy children participated in the study. The participants were divided equally by gender: 38 boys (47.5%) and 42 girls (52.5%). The mean age of the participants was 8.83 ± 1.55 years.

Measurement methods

Two bone age assessment methods were employed:

GP Bio Solution:^6,11 This software-based method uses body composition data derived from bioelectrical impedance analysis (BIA). Parameters such as height, weight, BMI, muscle mass, fat mass, and other body metrics are input into the system, and the software predicts bone age based on AI algorithms.

TW3 Method:³ The traditional method, where X-ray images of the left hand and wrist are analyzed according to the TW3 atlas. Bone age is determined by comparing the images to the standard skeletal maturity reference images.

Prediction model

The height prediction model in this study was developed using light gradient boosting machine with stratified sampling and 5-fold cross-validation for hyperparameter optimization. Separate models were built for males and females to capture sex-specific growth patterns. Model interpretation was performed using accumulated local effects to visualize how feature variations affect predictions and Shapley additive explanations to quantify individual feature contributions.

BIA measurements

Body composition was assessed using BIA by trained examiners who adhered to the standard protocols outlined by InBody Co., Ltd, as detailed in a previous report¹¹. These protocols required the exclusion of participants unable to maintain the prescribed posture, a minimum standing period of 5 min, a fasting interval of at least 2 h post-meal, and measurement in a post-void state. Additionally, potential interference during electrode placement was minimized by avoiding any extraneous physical contact or foreign materials.

Outcome measures

The primary outcome was to compare the bone age predictions from the two methods. The secondary outcomes included gender-based comparisons of bone age predictions, as well as exploring the correlation between body composition and bone age.

Statistical analysis

The primary analysis focused on comparing the differences in bone age predictions. According to previous studies^14,15, a non-inferiority margin of 0.661 years was established. This clinical trial compared growth age from the AI-based GP Bio Solution with bone age from the TW3 method using X-rays. The study was designed as an equivalence trial with the hypothesis H0: |µT−µC| ≥ δ versus H1: |µT−µC| < δ, where µT is the growth age from GP Bio Solution and µC is the bone age from TW3. The non-inferiority threshold, δ = 0.661 years, was adopted from prior study¹⁵ on VUNOmed-BoneAge, ensuring that differences between the methods remain clinically acceptable. The results were deemed equivalent if the 95% confidence interval for the difference in bone age remained within this margin. Secondary analyses included gender-based comparisons and exploratory analyses of body composition’s impact on bone age predictions.

Results

Demographic characteristics

A total of 80 children participated in the study, consisting of 38 males (47.5%) and 42 females (52.5%). The mean age of the participants was 8.83 ± 1.55 years. The baseline body composition measurements were as follows: average height of 136.99 ± 10.88 cm, weight of 34.43 ± 9.86 kg, and BMI of 18.03 ± 3.00 kg/m². Other body composition metrics included muscle mass (24.13 ± 5.84 kg), fat mass (8.73 ± 4.66 kg), and basal metabolic rate (925.05 ± 133.87 kcal). The participants were healthy, with no significant health conditions that would interfere with growth, based on the inclusion and exclusion criteria (Table 1).

Table 1.

Baseline demographic and body composition characteristics of study participants.

	N	Mean	Median	SD	95% CI
Age (years)	80	8.83	8.50	1.55	(8.48, 9.17)
Height (cm)	80	136.99	134.40	10.88	(134.57, 139.41)
Weight (kg)	80	34.43	32.50	9.86	(32.24, 36.62)
BMI (kg/m2)	80	18.03	17.60	3.00	(17.36, 18.69)
Protein (kg)	80	5.00	4.80	1.22	(4.73, 5.28)
Muscle (kg)	80	24.13	23.20	5.84	(22.82, 25.43)
Inorganid (kg)	80	1.88	1.79	0.45	(1.79, 1.98)
Fat (kg)	80	8.73	7.80	4.66	(7.70, 9.77)
SM (kg)	80	13.12	12.55	3.67	(12.30, 13.93)
BMR (kcal)	80	925.05	902.50	133.87	(895.26, 954.84)

Notes: BMI = body mass index; SM = skeletal muscle mass; BMR = basal metabolic rate; N = number of participants; SD = standard deviation; CI = confidence interval.

Primary efficacy outcome

The primary outcome was to evaluate the clinical equivalence of bone age predictions between GP Solution and the traditional X-ray-based TW3 method. The mean difference in bone age between the two methods was 0.04 ± 1.02 years. The 95% confidence interval (CI) for the difference in bone age ranged from − 0.18 years to 0.27 years, which falls within the pre-specified non-inferiority margin of 0.661 years. This result confirms that the bone age predictions made by the GP Solution are clinically equivalent to those made by the TW3 method (Table 2).

Table 2.

Comparison of bone age assessment between GP solution and TW3 method: overall and gender-specific analysis.

	N	Mean	Median	SD	95% CI
(a) Total
Treatment group (GP)	80	9.34	8.79	1.66	(8.97, 9.71)
Control group (TW3)	80	9.30	8.75	2.16	(8.82, 9.78)
Between-group differences	80	0.04	− 0.02	1.02	(− 0.18, 0.27)
(b) Male
Treatment group (GP)	38	9.45	8.75	1.83	(8.85, 10.05)
Control group (TW3)	38	9.34	8.38	2.28	(8.60, 10.09)
Between-group differences	38	0.11	0.23	0.99	(− 0.22, 0.43)
(c) Female
Treatment group (GP)	42	9.25	8.88	1.50	(8.78, 9.72)
Control group (TW3)	42	9.26	9.52	2.07	(8.62, 9.90)
Between-group differences	42	-0.01	-0.09	1.05	(− 0.34, 0.31)

Notes: GP = GP Solution (AI-based body composition method); TW3 = Tanner-Whitehouse 3 method; Between-group differences = GP minus TW3; N = number of participants; SD = standard deviation; CI = confidence interval. Non-inferiority margin was set at 0.661 years.

The analysis showed that the bone age difference between the two methods was minimal, with a slight advantage for the TW3 method in some individual cases, but overall, the GP Solution was able to predict bone age with comparable accuracy. This suggests that the software can be used as a reliable alternative to traditional radiographic methods, especially considering the non-invasive nature of the process and the reduced need for radiological exposure.

Secondary efficacy outcomes

Male Participants: In the male cohort, the mean bone age predicted by GP Solution was 9.45 ± 1.83 years, while the TW3 method predicted a mean bone age of 9.34 ± 2.28 years. The difference between the two methods was 0.11 ± 0.99 years, and the 95% confidence interval (CI) for the difference was − 0.22 years to 0.43 years. The results fell within the non-inferiority margin of 0.661 years, supporting the clinical equivalence of the two methods for males.

Female Participants: In the female cohort, the mean bone age predicted by GP Solution was 9.25 ± 1.50 years, while the TW3 method predicted a mean bone age of 9.26 ± 2.07 years. The difference in bone age was − 0.01 ± 1.05 years, with a 95% CI ranging from − 0.34 years to 0.31 years. This result also fell within the non-inferiority margin, confirming equivalence between the two methods for females. These findings suggest that the GP Solution performs similarly to the TW3 method in both sexes.

Exploratory analysis

Exploratory analyses demonstrated strong correlations between GP Solution bone age predictions and body composition parameters, particularly with height (r = 0.910), lean mass indicators (muscle: r = 0.883; protein: r = 0.877), and BMI (r = 0.463). Children with higher lean mass values showed correspondingly higher predicted bone ages, suggesting that the GP Solution effectively captures the established relationship between lean mass and skeletal maturity (Table 3).

Table 3.

Correlation analysis between body composition parameters and GP solution bone age assessment.

	Height	Weight	BMI	Protein	Muscle	Inorganic	Fat	SM	BMR
All	0.910	0.809	0.463	0.877	0.883	0.895	0.533	0.880	0.886
Male	0.891	0.807	0.506	0.892	0.896	0.888	0.524	0.895	0.897
Female	0.952	0.856	0.400	0.927	0.933	0.930	0.545	0.932	0.936

Note: Bold values indicate statistical significance (p < 0.05). BMI = body mass index; SM = skeletal muscle mass; BMR = basal metabolic rate; TW3 = Tanner-Whitehouse 3 method; GP = GP Bio Solution. Analysis performed for all participants (N = 80) and by gender subgroups (male: n = 38, female: n = 42).

When analyzing the difference between chronological age and predicted bone age, positive correlations were observed with body composition parameters, notably BMI (r = 0.359) and lean mass metrics. This relationship indicates that children with more advanced physical development relative to their chronological age demonstrated higher body composition values, particularly in parameters associated with maturation (Table 4).

Table 4.

Correlation analysis between body composition parameters and GP solution age and chronological age difference.

	Height	Weight	BMI	Protein	Muscle	Inorganic	Fat	SM	BMR
All	0.224	0.330	0.359	0.247	0.251	0.248	0.364	0.247	0.251
Male	0.470	0.552	0.475	0.523	0.522	0.497	0.477	0.518	0.522
Female	0.267	0.521	0.631	0.412	0.410	0.293	0.544	0.413	0.406

Note: Bold values indicate statistical significance (p < 0.05). BMI = body mass index; SM = skeletal muscle mass; BMR = basal metabolic rate; TW3 = Tanner-Whitehouse 3 method; GP = GP Bio Solution. Analysis performed for all participants (N = 80) and by gender subgroups (male: n = 38, female: n = 42).

In summary, these correlation patterns indicate that GP Solution bone age predictions align well with physiological maturation markers, providing additional validation for body composition-based skeletal maturity assessment. Moreover, the GP Solution method demonstrated clinical equivalence to the traditional TW3 method for bone age prediction, suggesting that body composition analysis could enhance growth assessments by providing complementary insights into children’s growth trajectories.

Safety and adverse events

No adverse events related to the use of X-ray or the GP Solution software were reported during the study. This suggests that the use of GP Solution as a non-invasive tool for bone age prediction is safe and well-tolerated, with no observed complications. The lack of adverse events further supports the potential of this software as a safe alternative to traditional methods that involve radiographic exposure. Overall, these results highlight the clinical equivalence of the GP Solution to the TW3 method and suggest that the AI-based software may offer a safer, faster, and more accessible option for bone age assessment in pediatric populations.

Discussion

Bone age assessment is a critical tool in pediatric endocrinology, particularly in evaluating growth potential and managing treatment for growth-related disorders¹⁶. The Tanner-Whitehouse 3 (TW3) method³, which relies on radiographic images of the hand and wrist, is currently considered the gold standard for bone age evaluation. However, the limitations of this method, such as radiation exposure, interobserver variability, and the need for specialized radiographic analysis, create challenges in everyday clinical practice. The development of an AI-based model using non-invasive body composition parameters to predict bone age could help overcome these limitations, thereby improving patient care and reducing the burden on healthcare resources. The primary objective of this study was to evaluate whether AH predictions using AI-based model derived from body composition biomarkers could be comparable to predictions made using traditional bone age in a cohort of normal Korean children. Our results demonstrate that there is no significant difference between the predictions based on body composition and those based on bone age, indicating that biological age can serve as a reliable predictor of AH. These findings support the potential use of body composition metrics as an effective, non-invasive alternative for AH prediction in clinical practice.

Growth is influenced by a complex interplay of genetic, nutritional, environmental, and hormonal factors¹⁷. Genetics determines a child’s growth potential, and parental height has a major impact on the final adult height¹⁸. However, environmental factors such as diet, physical activity, and general health status significantly influence how this genetic potential is realized¹⁹. Nutrition, in particular, has a direct impact on body composition by influencing fat mass, lean mass, and bone mineral density, which in turn affects skeletal growth. Adequate intake of essential nutrients such as proteins, vitamins, and minerals supports skeletal growth and muscle development, whereas deficiencies may lead to growth delays and altered puberty timing. Conversely, excessive fat accumulation can negatively affect growth through hormonal imbalances that interfere with growth hormone and insulin-like growth factor 1 (IGF-1) pathways, which are crucial for growth regulation¹⁸. Hence, assessing body composition provides a broader picture of a child’s growth trajectory, as it integrates the effects of both genetic and environmental influences^{7,8,10,20–23}. One significant finding of our study is that predictions based on biological age derived from body composition metrics were comparable to those from the TW3 method across both male and female participants. This finding has important clinical implications, especially in contexts where frequent exposure to radiation is a concern. Bone age assessments, typically performed during rapid growth periods such as early childhood or puberty, involve repeated radiographic imaging, which may increase cumulative radiation exposure. Therefore, utilizing body composition data collected through non-invasive methods like BIA represents a safer and more practical alternative for monitoring skeletal maturation²⁴ and predicting adult height.

The traditional method of using hand radiographs for bone age assessment has been a cornerstone in pediatric growth evaluation^2,3,25. However, there are notable limitations to this method that must be addressed. One such limitation is the interobserver and intraobserver variability in assessing bone age from radiographs. This variability is particularly prominent during late puberty, when skeletal maturation rates slow down, making it more difficult to precisely estimate bone age. Additionally, the variability tends to be higher during the assessment of certain carpal bones where subtle changes in growth plates during puberty can be challenging to discern accurately. Some studies have suggested that assessing bone age using radiographs of other skeletal regions, such as the elbow, may provide greater reliability during certain stages of growth, although this approach is not yet widely adopted in routine clinical practice^5,26. In cases where bone health is compromised, such as children with a history of fractures, or in conditions like rickets or osteogenesis imperfecta, traditional radiographs may not provide a clear indication of skeletal maturity²⁷. In such scenarios, the use of biological age based on body composition parameters can offer an alternative or complementary assessment method. In this study, we used an AI-based model that incorporates body composition parameters such as FFM, fat mass, BMI, muscle mass, and other related metrics to predict adult height^{7,10,23,24,28}. These parameters reflect not only the genetic potential but also the child’s current nutritional and physical activity status. The integration of these metrics into the growth assessment process adds a layer of dynamism that is not captured by bone age alone, which primarily focuses on the skeletal aspect of growth. Our study suggests that fat-free mass, in particular, serves as a reliable biomarker for predicting adult height, as it is closely linked to skeletal muscle and bone development. FFM is also indicative of nutritional status, which plays a significant role in growth potential, particularly during puberty^8,13. The use of body composition metrics, therefore, provides a more comprehensive understanding of growth potential, emphasizing the combined impact of both genetic and environmental factors.

Another unique aspect of our study is the application of body composition to predict adult height specifically in Korean children. While several international studies have examined the relationship between body composition and growth in different populations, relatively few have focused on Korean children. Our findings suggest that body composition could serve as a practical and effective predictor of adult height in this population, potentially providing valuable insights into growth patterns that may differ due to genetic, dietary, and cultural factors. Over recent decades, significant changes in lifestyle, including diet and physical activity levels, have influenced body composition trends among Korean children²⁹. The use of non-invasive methods such as BIA in this context is particularly valuable, as it allows for more widespread application without the risks associated with repeated radiographic exposure. This approach aligns with the increasing emphasis on personalized, data-driven healthcare, where individualized assessment of growth potential can lead to more tailored and effective interventions.

Despite the promising results, it is important to acknowledge the limitations of our study. First, the study was conducted on a relatively homogenous population of healthy Korean children, which may limit the generalizability of the findings to other populations or to children with specific growth disorders. The inclusion of children with growth hormone deficiencies, precocious puberty, or other endocrine disorders in future studies would help validate the applicability of our findings in more diverse clinical settings. Second, although the AI model based on body composition showed equivalence to the traditional bone age method, additional refinement of the model is needed to enhance its predictive accuracy across different growth phases. Longitudinal studies tracking growth from early childhood through puberty and into late adolescence could provide valuable insights into the long-term predictive power of body composition metrics and help fine-tune the model for different growth stages. Another important consideration is the potential variability in body composition measurements obtained through BIA³⁰. Factors such as hydration status, time of day, and physical activity can affect BIA measurements, potentially introducing variability in the results²². Ensuring that standard protocols are followed for BIA measurement, such as consistent timing and pre-measurement conditions, is essential for minimizing variability and improving the reliability of the results. Therefore, although we adopted a standardized measurement protocol to mitigate the inherent limitations of BIA, we acknowledge that certain challenges persist. Specifically, we could not ensure uniform hydration status or electrolyte balance across all participants, and body temperature variations were not recorded. Even though our protocol aligns with established clinical practices, these factors may still introduce variability in the BIA measurements. Furthermore, advancements in BIA technology and AI algorithms could further enhance the precision and utility of body composition-based predictions. Moreover, to further enhance and validate the predictive accuracy of our AI-based model, future research should include a larger population. Increasing the sample size would improve the model’s robustness, allow for better control of confounding factors, and ensure broader applicability across diverse pediatric populations. Despite these limitations, our study has several notable strengths. The prospective design allowed for comprehensive data collection, including both bone age and body composition metrics, which are crucial for understanding the relationship between skeletal and body composition development. The use of an independent validation cohort also strengthens the reliability of our findings, demonstrating that body composition-based predictions are consistent with bone age predictions even when applied to a different group of children. This consistency underscores the robustness of our AI-based model and its potential utility in clinical practice. Importantly, the use of non-invasive, radiation-free methods to assess growth potential aligns with the principle of minimizing harm, particularly in pediatric populations.

In conclusion, this study demonstrates that body composition biomarkers provide a viable and clinically equivalent alternative to bone age for predicting adult height in normal Korean children. The use of AI to integrate metrics such as FFM, BMI, and fat mass provides a comprehensive and individualized growth assessment, reducing the reliance on radiographic imaging and associated radiation exposure. This non-invasive approach not only complements traditional bone age assessments but also provides additional insights into the role of nutrition and physical activity in growth. Future research should focus on validating this approach in a broader range of populations, including children with growth disorders, and on further refining the predictive models to enhance accuracy. By integrating body composition assessments into routine pediatric care, we can move towards a more holistic, personalized approach to growth monitoring, ultimately improving outcomes for children at risk of growth abnormalities.

Author contributions

H.W.J., D.C., J.H.C., K.L. acquisition of data, and analysis. J.H.C., J.H.L., J.K., W.Y.J. conception, desige, and interpretation of data. H.W.J., D.C. drafting the article. H.W.J. J.K., W.Y.J. final approval of article.

Funding

This work was supported by the Korea Medical Device Development Fund granted by the Korean government (Ministry of Science and ICT, Ministry of Trade, Industry and Energy, Ministry of Health & Welfare, and Ministry of Food and Drug Safety) (KMDF_PR_20200901_0039, KMDF_PR_20200901_0293, RS_2023_00243310). This work was supported by a National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (NRF-2022R1A2C2092726), Korea University Anam Hospital, Seoul, Republic of Korea (Grant No. K2305161, K2313001, K2312991, I2300231, I2203971).

Data availability

The data that support the findings of this study are available from author, Woo Young Jang but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Furthermore, all data generated or analyzed during the current study will not be disclosed due to policy of the Korea University Anam Hospital Research Ethics Board.

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

This study was approved by the Institutional Review Board (IRB) of Korea University Anam Hospital and Kyung Hee University Medical Center, and informed consent was obtained from both the parents and the pediatric participants.