Artificial intelligence for pediatric height prediction using large-scale longitudinal body composition data

Abstract

Objective

We developed a precise, reliable artificial intelligence (AI) model for predicting the future height of children and adolescents based on anthropometric and body composition data.

Materials and Methods

We used an extensive longitudinal dataset from a large-scale Korean cohort study, which included 588,546 measurements from 96,485 children and adolescents aged 7–18. We developed a prediction model using the light gradient boosting method and integrated anthropometric and body composition metrics along with their standard deviation scores (SDSs) and velocity parameters. Model performance was assessed through root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). We employed Shapley additive explanations (SHAP) for model interpretability.

Results

The model accurately predicted future heights. For males, the average RMSE, MAE, and MAPE were 2.51 cm, 1.74 cm, and 1.14%, respectively, with female prediction results showing comparable accuracy (2.28 cm, 1.68 cm, and 1.13%, respectively). Shapley additive explanations analysis revealed that the SDS of height, height velocity, and soft lean mass velocity were key predictors of future height. The model created personalized growth curves through estimation of individual-specific height trajectories, comparison with actual measurements, and identification of key variables using local SHAP values.

Conclusion

Our model produces accurate and personalized growth curves, incorporating explainable AI techniques for enhanced clinical understanding. This method advances pediatric growth assessment and provides robust clinical decision support. Despite limitations including the absence of handwrist radiography comparison and Korean population specificity, our approach demonstrates significant potential for early identification of growth disorders and optimization of growth outcomes.

Introduction

Evaluating growth and predicting height in children and adolescents are essential parts of pediatric care. Linear growth serves as a crucial health indicator that reflects the cumulative impact of genetic, environmental, and socioeconomic factors.^1–3 Tracking linear growth allows for the early identification of growth disorders such as growth hormone deficiency, celiac disease, and chronic conditions affecting development.^4–7 Predicting future height accurately is crucial for diagnosing growth disorders, starting hormone therapy, and assessing treatment effectiveness.^8–10 Traditional height prediction methods rely on skeletal maturity assessment using handwrist radiographs. These include the Bayley–Pinneau, ¹¹ Tanner–Whitehouse, ¹² and Roche–Wainer–Thissen ¹³ methods. Nonetheless, these methods have drawbacks, including radiation exposure, a requirement for specialized expertise, and high interobserver variability.^14–16 Additionally, their accuracy diminishes in children with growth disorders or those receiving growth hormone therapy.^17–19

Considering the limitations of conventional growth assessment techniques, body composition measurements have become vital complementary indicators for tracking and forecasting pediatric growth and development.^20–25 Recent advancements in body composition assessment methods, such as multifrequency bioelectrical impedance analysis (BIA), have facilitated the collection of large-scale and extended pediatric data.^26–30 Machine learning (ML) techniques are suitable for analyzing these large-scale longitudinal datasets in growth modeling and height prediction. In growth-related fields, ML has been primarily applied to bone age assessment through image processing. ^{15 31–35} Several studies have directly predicted height using ML; however, they relied only on basic anthropometric data and did not consider body composition measurements.^36,37 The recent emergence of extensive body composition databases has fueled innovative research, including models for height estimation using ML ³⁸ and biological maturity evaluation frameworks that complement traditional skeletal maturity methods. ³⁹

This study addressed the existing gap in predicting pediatric growth by utilizing ML algorithms on a large-scale, longitudinal dataset that integrated anthropometric and body composition measures. The gradient boosting framework effectively captured complex relationships between growth-related factors and height trajectories through iterative pattern learning. Such methodology facilitated the development of precise predictive models that identified critical growth determinants. The framework accommodated predictions from multiple baseline ages with varying prediction horizons, enabling growth trajectory modeling throughout pediatric development.

Our method presented several distinctive benefits. First, we integrated the standard deviation scores (SDSs) for body composition measures, which were obtained from a comprehensive body composition database. ⁴⁰ These SDSs measured growth status in age- and sex-matched groups, providing standardized comparisons that went beyond absolute anthropometric and body composition values. Our longitudinal data additionally allowed us to calculate growth velocities, which are often overlooked in cross-sectional studies. Tracking growth velocities in height and body composition is vital for detecting abnormalities in growth patterns.^41–44 Incorporating body composition velocity provided valuable insights into the dynamic physical development of children, enabling a more comprehensive evaluation of growth trajectories. Additionally, we improved model interpretability by implementing explainable artificial intelligence (AI) techniques, such as the use of Shapley additive explanations (SHAP), to examine the key factors influencing growth predictions.^45–50 Our model integrated personalized growth curve estimation with local interpretability. This combination facilitated the early identification of at-risk children and enabled targeted interventions in conjunction with conventional growth monitoring techniques.

The structure of this paper is as follows. Section “Data and methods” outlines the data and methods used. Section “Results” highlights the results from our height prediction model and the analyses related to explainable AI. Section “Discussion” presents our findings and their implications, while Section “Conclusion” concludes the paper and proposes directions for future research.

Data and methods

Study population

The AI model was developed using data from the GP Co., Ltd Cohort Study (hereafter, GP Cohort Study), a long-term research project initially based in Gwangmyeong, South Korea. Gwangmyeong is an urban area adjacent to Seoul in Gyeonggi Province, which encompasses 44.8% of Korea's total population. Since 2015, data collection has expanded to include additional urban and suburban areas, including Osan, Siheung, and Ansan in Gyeonggi Province, as well as selected regions in other provinces. This study encompassed students from elementary, middle, and high schools aged 7–18, with an average participation of 35 schools each year. The cohort was not restricted to healthy children but included the general school population. Data collection began on 1 January 2013, and accumulated 588,546 measurements from 96,485 children and adolescents (50,480 males and 46,005 females) born between 1998 and 2016. Inclusion criteria comprised: (1) ability to maintain standard posture for anthropometric measurements and (2) written informed consent from guardians with assent from participants aged 12 years or older. Exclusion criteria included: (1) inability to undergo standard measurement procedures due to physical limitations, (2) absence of informed consent, and (3) noncompliance with premeasurement protocols such as the 2-h fasting requirement for body composition analysis. The Institutional Review Board of Korea University Anam Hospital approved the study protocol (2025AN0110). Due to the retrospective design with de-identified data, the IRB waived the informed consent requirement for data analysis.

Data collection and measurements

Measurements were conducted by dedicated staff, from GP Co., Ltd, who have performed standardized assessments since 2013. All measurements followed written protocols based on CDC anthropometric guidelines and manufacturer specifications. Heights were measured using a stadiometer, while body composition was assessed using octopolar multifrequency BIA (InBody models J10 and J30, InBody Inc., Seoul, Korea). The measurement staff were not involved in the subsequent data analysis or interpretation.

The predictor variables of interest comprised height, weight, protein mass, soft lean mass (SLM), body fat mass (BFM), skeletal muscle mass (SMM), bone mineral mass, total body water (TBW), basal metabolic rate (BMR), as well as waist and hip circumferences. The following indices were calculated: body mass index (BMI), BFM index, SMM index, and waist-to-hip ratio (WHR):

$$\mathrm{BMI}(\mathrm{kg}/{\mathrm{m}}^2)={\displaystyle \frac{\mathrm{Weight}(\mathrm{kg})}{\mathrm{Height}{(\mathrm{m})}^2},}$$

$$\mathrm{BFMI}(\mathrm{kg}/{\mathrm{m}}^2)={\displaystyle \frac{\mathrm{BFM}(\mathrm{kg})}{\mathrm{Height}{(\mathrm{m})}^2},}$$

$$\mathrm{SMMI}(\mathrm{kg}/{\mathrm{m}}^2)={\displaystyle \frac{\mathrm{SMM}(\mathrm{kg})}{\mathrm{Height}{(\mathrm{m})}^2},}$$

$$\mathrm{WHR}={\displaystyle \frac{\mathrm{Waist}\mathrm{circumference}(\mathrm{cm})}{\mathrm{Hip}\mathrm{circumference}(\mathrm{cm})}.}$$

The dataset preprocessing involved the following procedures. Observations containing missing values were removed at the observation level. Age was calculated in months from birth to measurement dates. ^a Weight measurements were validated against the sum of body composition components (SLM, bone mineral mass, and BFM), with excessive discrepancies excluded to ensure data quality and measurement reliability. Statistical outlier detection was performed separately by sex and age group using the interquartile range (IQR) method. The anthropometric measurements of children and adolescents were standardized using the 2017 Korean National Growth Charts. This included height, weight, and BMI. Body composition parameters not available in these charts were standardized using the GP growth chart developed by Chun et al.^{40, b} The dataset included both original measurements and their corresponding age- and sex-specific SDS. The preprocessed dataset contained 27 attributes, which included fundamental variables such as sex and age, as well as metrics related to anthropometry and body composition measures and associated SDS.

Data preparation

The target of the prediction was set as the future growth rate in height, determined through the following method. Constructing the prediction dataset involved selecting paired observations from participants with multiple measurements. For each participant, we selected two observations: the earlier observation (baseline) and the later observation (target). The height from the later observation (Height_T) was used to calculate the prediction target, with the age at the target time point (Months_T) included as a predictor to account for the prediction horizon. The growth rate, calculated as

$$\mathrm{Growth}\mathrm{\_}\mathrm{Rate}={\displaystyle \frac{\mathrm{Height}\mathrm{\_}\mathrm{T}-\mathrm{Height}\mathrm{\_}\mathrm{baseline}}{\mathrm{Height}\mathrm{\_}\mathrm{baseline}}\times 100(\text{\%}),}$$

served as the target variable for baseline observation. ^c Figure 1 illustrates this data structure. This method required participants to have at least two measurements for longitudinal analysis.

Figure 1.

Example of a prediction dataset with target variables. Note: The figure illustrates an example of a prediction dataset, showing the inclusion of the target variable (Growth_Rate) alongside predictor variables. The age in months at the target observation (Months_T) is incorporated as a predictor variable for the baseline observation. This process is applied to all entities with two or more observations, creating a prediction dataset.

Prediction horizons between paired observations varied across participants (median: 19 months, IQR: 9–34 months). After applying preprocessing procedures to remove observations with missing values, measurement errors, and statistical outliers, then restricting the analysis to ages 7–16 years and excluding participants with single measurements, the final dataset comprised 44,683 participants (23,402 males and 21,281 females) with sufficient longitudinal data, generating 206,112 training samples (106,461 for males and 99,651 for females).

Additional predictor variables were included to effectively capture growth dynamics and trajectories. Growth_Rate_BM represents the expected growth rate if the participant maintains their current height SDS and serves as a benchmark for comparison. Growth velocity variables were calculated from historical measurements taken 3–6 months before baseline. These included height velocity (Height_V) and body composition velocities (weight velocity [Weight_V], Protein_V, SLM velocity [SLM_V], BFM_V, SMM_V). The velocities represent the rate of change per month and provide information about recent growth patterns that predict future trajectories.

Growth velocity calculations required measurements from 3 to 6 months before baseline. Of the 44,683 participants in the prediction dataset, 17,948 were excluded due to insufficient prior measurements. The remaining 26,735 participants (13,907 males and 12,828 females) generated 105,774 measurements (54,463 from males and 51,311 from females). The model's final input dataset was created by combining the target variable (Growth_Rate) with the prediction target information (Months_T, Growth_Rate_BM, and Height_BM) for each subject. This final dataset comprises 37 columns, encompassing the initial 27 attributes, 6 growth velocity metrics, 3 prediction target variables, and the target variable. Of these, 35 columns served as predictor variables, excluding the target variable and sex, as separate models were trained for males and females. Supplementary Table S1 provides details on the dataset's variables, including their descriptions and data types.

1. Prediction Dataset Creation	206,112 measurements (44,683 subjects) 106,461 males, 99,651 females
Preprocessing the original measurements Identify multiple observations per subject
Select two observations per subject (baseline and target)
Calculate the target variables (Growth_Rate)
Construct a dataset with the target variables
2. Feature Engineering	105,774 measurements (26,735 subjects) 54,463 males, 51,311 females Thirty-seven columns (35 predictors)
Incorporation of benchmark features (Height_BM, Growth_Rate_BM)
Computation of the growth velocity variables
3. Data Integration and Transformation
Merge target variables and prediction information
Normalize the target variable (Yeo-Johnson method)
Scale Predictor Variables (Robust Scaler)

Prediction model

The height prediction model was trained using the Light Gradient Boosting Method, a framework that utilizes tree-based learning algorithms for gradient boosting. This framework is well-known for its efficiency and accuracy when handling large datasets. ^d The dataset was divided using a stratified sampling method based on individual identification numbers. A random 20% of individuals formed the test set, while the remaining 80% constituted the training set. This approach ensured that data from the same individual did not overlap between the training and test sets, thereby avoiding data leakage and reducing the risk of overestimating model accuracy. Distinct models were created for males and females to account for possible sex-specific variations in growth patterns. The representativeness and comparability of the training and test datasets were evaluated by comparing the distributions of key variables and calculating the median and IQR for each variable.

Hyperparameter optimization was conducted through grid search, accompanied by 5-fold cross-validation on the training dataset. The training set was divided into five equal-sized folds, with the model trained on four of these and validated on the remaining fold. This cycle was repeated five times, ensuring each fold acted as the validation set once. The optimized hyperparameters comprised the maximum depth of the trees, the maximum number of leaves per tree, the number of boosting rounds, and the learning rate. These structural parameters serve as regularization mechanisms that constrain model complexity to prevent overfitting. The best hyperparameter configuration from the cross-validation was applied to train the final model on the entire training set. The trained model was then evaluated on an independent test set.

We assessed the performance of the trained models using the test set, calculating three common performance metrics: root mean squared error (RMSE), mean absolute percentage error (MAPE), and mean absolute error (MAE). To evaluate the model's stability and reliability, we employed a bootstrapping method, randomly dividing the test set into 50 subsets. We computed the performance metrics (RMSE, MAE, and MAPE) for each subgroup and estimated the standard deviation and 95% confidence intervals (CIs) for each metric. Heatmaps illustrating MAPE values across different age groups and prediction horizons were created to examine the model's performance. Figure 2 illustrates the flowchart of the model development pipeline, from data preparation to model evaluation.

Figure 2.

Model development pipeline overview.

Explainable AI

The model for predicting height was analyzed using three explainable AI techniques: feature importance, accumulated local effects (ALEs), and SHAP. These methods highlighted the key factors impacting the predicted heights of children and adolescents. Feature importance measures the relative contribution of each feature to the model's predictions. The importance values were determined based on the average gain of each feature, with higher values indicating a greater influence. To facilitate interpretation and comparison, these values were normalized to a scale of 100. Accumulated local effect plots display the average impact of individual features on model predictions across their respective value ranges, considering interactions between features. These plots illustrate the cumulative local effect on predicted growth in height as the value of a specific feature changes while keeping other features constant. The analysis focused on the key features identified by their importance scores. Shapley additive explanations provided a further method to elucidate the impact of features on model predictions. Using game-theoretic Shapley values, this method assesses the marginal impact of each feature on all projections, enabling an analysis of both overall importance and individual cases. Shapley additive explanations values were derived for every instance and visualized through summary plots.

Personalized growth curve estimation

Estimating personalized growth curves is a crucial aspect of our methodology, enabling individualized assessments of growth. This method produces a tailored growth trajectory by forecasting heights at designated ages. The predicted curves are displayed alongside actual height measurements and population-level benchmarks. We utilized local SHAP to enhance model interpretability by pinpointing and emphasizing the most notable features influencing individual predictions. To illustrate these concepts, we present case studies of randomly chosen male and female subjects in the results section.

Results

Baseline characteristics

Table 1 summarizes the baseline characteristics of the study population, categorized by sex and dataset (training vs. test). In the case of male participants, the training set comprised 261,567 data points from 11,125 unique individuals, while the test set contained 66,840 data points from 2782 unique males. For female participants, the training set included 247,580 data points from 10,262 unique individuals, and the test set had 65,104 data points from 2566 unique females. The table shows the median and IQR for essential variables, including age (in years), height, weight, BMI, protein mass, SMM, BMR, TBW, and the target height.

Table 1.

Baseline characteristics.

	Male		Female
	Training set (n = 261,567)	Test set (n = 66,840)	Training set (n = 247,580)	Test set (n = 65,104)
Number of Entities	11,125	2782	10,262	2566
Age (months)	9.08 [8.08–10.50]	9.00 [8.17–10.42]	9.00 [8.08–10.33]	8.92 [8.08–10.33]
Height	134.9 [128.6–142.8]	135.0 [128.9–142.8]	133.6 [127.3–142.5]	133.2 [127.2–142.0]
Weight	32.3 [27.2–40.3]	32.3 [27.1–39.9]	30.6 [25.9–37.5]	29.8 [25.5–36.6]
BMI	17.6 [16.0–20.2]	17.5 [15.9–20.0]	17.0 [15.5–19.1]	16.7 [15.3–18.7]
Protein	5.0 [4.4–5.8]	5.0 [4.4–5.8]	4.6 [4.1–5.5]	4.6 [4.1–5.4]
SLM	23.9 [21.1–27.9]	23.9 [21.2–27.9]	22.2 [19.7–26.2]	21.9 [19.5–25.7]
BFM	6.3 [4.0–11.0]	6.0 [4.0–10.0]	7.0 [5.0–10.0]	6.0 [4.0–10.0]
SMM	13.0 [11.3–15.6]	13.0 [11.3–15.5]	12.0 [10.4–14.4]	11.8 [10.3–14.2]
BMR	917 [853–1009]	917 [855–1009]	879 [821–971]	873.0 [817–960]
TBW	18.6 [16.4–21.7]	18.6 [16.5–21.7]	17.3 [15.3–20.4]	17.1 [15.2–20.0]
Target Height	147.8 [138.8–159.5]	147.9 [138.8–158.9]	148.3 [138.3–156.3]	148.0 [138.1–156.0]

The table presents the median and interquartile range [1st quartile–3rd quartile] for variables in the training and test sets. The number of data points (n) and unique individuals (Number of Entities) are also reported for each set.

BFM: body fat mass; BMI: body mass index; BMR: basal metabolic rate; SLM: soft lean mass; SMM: skeletal muscle mass; TBW: total body water.

In the training set, males had a median age of 9.08 years, a height of 134.9 cm, weight of 32.3 kg, BMI of 17.6 kg/m², protein mass of 5.0 kg, SLM of 23.9 kg, BFM of 6.3 kg, SMM of 13.0 kg, BMR of 917 kcal, TBW of 18.6 kg, and a target height of 147.8 cm. For females in the training set, the median age was 9.00 years, with a height of 133.6 cm, weight of 30.6 kg, BMI of 17.0 kg/m², protein mass of 4.6 kg, SLM of 22.2 kg, BFM of 7.0 kg, SMM of 12.0 kg, BMR of 879 kcal, TBW of 17.3 kg, and a target height of 148.3 cm. The test set displayed comparable values for both sexes, demonstrating a suitable balance between the training and test sets.

Prediction accuracy

Table 2 displays the prediction accuracy metrics, indicating high accuracy for both male and female models. For males, the models yielded a mean RMSE of 2.51 ± 0.07 cm [95% CI 2.37–2.66 cm], MAE of 1.74 ± 0.05 cm [95% CI 1.65–1.83 cm], and MAPE of 1.14% ± 0.03% [95% CI 1.08%–1.20%]. In contrast, females displayed a mean RMSE of 2.28 ± 0.07 [95% CI 2.15–2.41] cm, MAE of 1.68 ± 0.05 [95% CI 1.58–1.78] cm, and MAPE of 1.13 ± 0.03 [95% CI 1.06–1.19].

Table 2.

Overall prediction accuracy.

Sex	RMSE	MAE	MAPE
Male	2.51 ± 0.07	1.74 ± 0.05	1.14 ± 0.03
	(2.37–2.66)	(1.65–1.83)	(1.08–1.20)
Female	2.28 ± 0.07	1.68 ± 0.05	1.13 ± 0.03
	(2.15–2.41)	(1.58–1.78)	(1.06–1.19)

The table presents the performance metrics for height prediction models, including RMSE, MAE, and MAPE. Data are presented as the mean ± SD (95% CI) for each metric, allowing for an assessment of the model's stability and reliability across 50 different test data subsets.

CI: confidence interval; MAE: mean absolute error; MAPE: mean absolute percentage error; RMSE: root mean square error.

To evaluate the performance of the prediction model, we created heatmaps that illustrate the MAPE across various age groups and prediction horizons for both males and females (Figure 3). These heatmaps indicated that prediction accuracy fluctuated based on the current age and length of the prediction horizon. Overall, the model displayed lower MAPE for shorter prediction horizons and older ages in both sexes. Mean absolute percentage error values rose with extended prediction horizons and younger ages, as shown by the lighter shades in the upper-left section of the heatmaps. Despite these fluctuations, the model exhibited strong performance, with the highest MAPE remaining under 2.42% for males (predicting 5 years ahead from age 7) and 2.00% for females (predicting 5.5 years ahead from age 10).

Figure 3.

Mean absolute percentage error (MAPE) heatmaps across different age groups and prediction horizons. Notes: Heatmaps depicting MAPE across different age groups and prediction horizons for (a) males and (b) females. The MAPE is calculated for each combination of age (in years) on the x-axis and prediction horizon (in years) on the y-axis. The color scale indicates the MAPE values, with darker shades of red representing lower errors.

Feature importance

We identified the most significant height predictors by calculating feature importance values, which represent the relative contribution of each feature to the prediction model based on its average gain (Figure 4). The top 20 features for each sex were visualized using horizontal bar plots, with importance values normalized to a scale of 100. We note that prediction target information, such as Height_BM, Months_T, Months, and Growth_Rate_BM, was excluded from this plot to focus on the anthropometric and body composition variables. The results showed that the SDS of current height (SDS_Height) had the greatest influence on both males and females, followed by Height_V and SLM_V after considering the prediction target. Beyond these variables, additional top predictors of males included height, WHR, and SMM. For females, other key predictors were WHR, Weight_V, and height.

Figure 4.

Feature importance plots.

Note: Importance of features within the height prediction models for (a) males and (b) females. The feature importance values are normalized to a total sum of 100, with higher values indicating greater importance. To focus on the importance of the anthropometric and body composition variables, we excluded prediction target information variables (Height_BM, Months_T, Months, and Growth_Rate_BM) from plots. Among these excluded variables, Growth_Rate_BM had an overwhelmingly high importance of 96.33% for males and 96.11% for females.

Accumulated local effect

The ALE plots illustrate the average influence of individual features on model predictions, considering feature interactions (Figure 5). We observe a strong negative and nonlinear association between the SDS_Height and the estimated growth in height. A distinct negative correlation was observed when the scaled SDS_Height was below 0, whereas no significant correlation was found for values above 0. Likewise, a negative association was noted between the Height_V and the predicted growth in height. Conversely, the SLM_V showed a positive correlation with the anticipated growth in height.

Figure 5.

Accumulated local effect (ALE) plots. Note: ALE plots for influential features in height prediction models for (a) males and (b) females, as determined by feature importance analysis. The figure illustrates the cumulative local effect on the predicted growth in height as each feature value changes, while holding other features constant. The x-axis displays the range of values for the selected feature, while the y-axis indicates the corresponding change in the predicted growth in height relative to the average prediction. Note that due to variable transformations, changes in the scaled target and predictor variables do not directly translate to changes in their original units.

Shapley additive explanations

The SHAP summary plots demonstrate the impact of features on model predictions by calculating Shapley values (Figure 6). Red represents features that increase predictions, whereas blue denotes those that decrease them. Important variables for both sexes include fundamental anthropometric metrics such as benchmark growth rate (Growth_Rate_BM) and SDS_Height. Growth velocities, including SLM_V and Height_V, had a significant effect on model results. Notably, the height SDS and velocity exhibited inverse relations with the predicted growth rate, while the SLM_V showed a positive correlation, consistent with previous findings.

Figure 6.

Shapley additive explanation (SHAP) summary plots. Note: SHAP summary plots of models for predicting height in (a) males and (b) females. The plots illustrate the overall impact of each feature on the model's predictions, with features increasing the prediction shown in red and those decreasing the prediction shown in blue. Features are listed on the y-axis in descending order of importance. The x-axis represents the SHAP values, indicating the magnitude and direction of each feature's impact. The variable “Months” represents the current age in months, while “Months_T” indicates the target age in months.

Individual height prediction curves

A major advantage of our height prediction model is its ability to create personalized growth curves for children and adolescents. ^e To highlight this feature and enhance local interpretability, we present two illustrative examples (Figure 7). These examples were randomly chosen from the test set, featuring one boy and one girl, referred to as Boy #1 and Girl #1, respectively.

Figure 7.

Personalized growth and Shapley additive explanation (SHAP) decision plots. Note: Personalized growth curves and SHAP decision plots for randomly selected (a) boy and (b) girl from the test set. The customized growth curves were generated using each child's longitudinal data. The SHAP decision plots quantify each feature's impact on the estimated growth in height, with positive values indicating an increase and negative values a decrease in the predicted growth rate. The variable “Months” represents the current age in months, while “Months_T” indicates the target age in months.

Figure 7 presents the predicted and actual growth curves along with the associated SHAP decision plots for Boy #1 and Girl #1. These growth curves were created using the trained model and the available longitudinal data for each child. For Boy #1, the Growth_Rate_BM (2.556) had the most significant positive impact on predicted growth in height. Additionally, the WHR (1.181) and Height_V (0.769) also contributed positively to the predicted growth in height. In contrast, for Girl #1, the Growth_Rate_BM (2.028) was the most notable positive contributor. The TBW percentage (TBWP; 0.864), the SDS of BFM (SDS_BFM; −0.909), and the SDS of SMM (SDS_SMM; −0.517) further illustrate positive influences on the predicted height.

Discussion

In this study, we developed an accurate and reliable AI model to predict the future heights of children and adolescents by analyzing extensive anthropometric and body composition data from a large-scale longitudinal research project. The prediction model yielded mean RMSEs of 2.51 cm for males and 2.28 cm for females, MAEs of 1.74 cm for males and 1.68 cm for females, along with MAPEs of 1.14% for males and 1.13% for females. Performance varied by age group and prediction horizon, showing improved accuracy for shorter prediction intervals and older ages. Importantly, the model maintained high accuracy across all scenarios, with MAPE values consistently below 2.42% for males and 2.00% for females in every age group and prediction horizon.

Previous height prediction studies demonstrate varied accuracies across methodologies. Direct ML prediction from anthropometric data yielded MAE of 3.35–3.66 cm ³⁶ and 2.46–3.00 cm. ³⁷ Deep learning methods using radiographic images showed MAE of 4.62 cm in Korean children. ³⁴ Traditional bone age-based methods produced RMSE of 2.7–3.3 cm with automated BoneXpert ³² and SD of 2.5–2.8 cm with Tanner–Whitehouse in athletic populations. ⁵¹ Bayley–Pinneau demonstrated SD of 3.9–4.2 cm in Korean short stature children. ⁵² In ISS cohorts, multiple linear regression models achieved SD of 3.70–4.19 cm while traditional Bayley–Pinneau showed SD of 5.81 cm. ⁵³ Previous investigations primarily analyzed limited sample sizes from restricted cohorts or populations with specific clinical characteristics. Our analysis examined 206,112 training samples from 44,683 Korean children aged 7–16 years from general school populations. Despite analyzing this heterogeneous population spanning normal growth patterns and diverse clinical conditions, our model achieved superior accuracy to established methods.

Our findings emphasize the crucial roles of body composition and growth velocity in predicting future height. Notably, the analysis of feature importance, along with the SHAP results, pinpointed height and SLM_V as the key variables. The ALE plots unveiled interesting connections between these growth velocities and anticipated height increases. A negative correlation between Height_V and expected growth in height supports the idea of catch-up growth, in which children who initially experience growth delays may have accelerated growth in later periods.^54–56 Conversely, the SLM_V exhibited a positive correlation with predicted growth in height, suggesting that children with greater muscle accumulation are likely to experience increased height. Nonetheless, the connection between lean mass increase and future height in children and adolescents is still a matter of debate. Although these results highlight the intricate relation between lean mass development and height achievement, the observational nature of the study limits our conclusions to possible correlations rather than establishing causal relations.

The generation of personalized growth curves is a notable feature of this model. The integration of explainable AI techniques, particularly SHAP decision plots, provides interpretable insights into the factors driving individual growth predictions. For example, the SHAP decision plot for Boy #1 shows that his predicted height was primarily influenced by his SDS of height, WHR, and Height_V. In contrast, Girl #1's predicted growth in height was mainly affected by her SDS of height, TBWP, and SDS_BFM. These personalized insights demonstrate the model's potential for tailored growth assessments.

Our study enhances pediatric growth assessment by applying ML algorithms to a comprehensive set of anthropometric and body composition measures. Conventional pediatric height predictions usually rely on skeletal maturity assessment using handwrist radiographs,^11–13 which have limitations such as the need for specialized expertise, high interobserver variability, and reduced accuracy in assessing children with growth disorders.^14–19 Our study provides complementary information by utilizing body composition data from BIA, which is noninvasive, cost-effective, and convenient, enabling large-scale and long-term data collection.^27–30 This method offers additional insights into biological maturity beyond skeletal maturity.^20–25

To effectively leverage longitudinal body composition data, our study employed SDS for measuring body composition and growth velocities. These SDS, derived from a comprehensive database, provide a standardized method to quantify relative growth status within age- and sex-matched cohorts. ⁴⁰ This standardization facilitates the identification of growth patterns across various populations and improves the comparability of growth data in longitudinal studies. Additionally, our longitudinal data enabled the calculation of growth velocities, which are crucial for detecting deviations from the typical growth pattern.^41–44 The inclusion of body composition velocity offers unique insights into children's physical development dynamics, deepening our understanding of growth trajectories.

This study employs ML techniques to analyze large biometric datasets, aiming to directly predict height in children and adolescents. It overcomes the constraints of past research, which primarily focused on utilizing radiographic data for automated skeletal maturity assessments. ^{15 31–35} Previous efforts to predict height directly were limited by smaller datasets and insufficient body composition measures.^36,37 This study resolves these issues by incorporating comprehensive body composition data from the extensive GP Cohort Study. By applying specialized big data methods, the model gathers insights from various variables, enabling the creation of an unbiased, objective, and reliable AI model that delivers consistent results.

The integration of SHAP analysis provides model transparency by revealing how body composition parameters influence height predictions. This interpretability identifies population-level growth determinants and offers clinicians’ evidence-based reasoning for individual predictions. Furthermore, SHAP analysis revealed distinct predictive patterns between sexes. Males relied primarily on SDS of height, SLM_V, Height_V, and SMM. Females utilized SDS of height with different body composition markers including SDS_SMM, WHR, and TBWP. Weight velocity emerged as a key predictor exclusively in female models. The identification of these distinct predictive patterns validates our sex-stratified modeling approach, given that males and females follow different growth trajectories during adolescence. ⁵⁷

For clinical implementation, our AI model enhances pediatric growth assessment through height trajectory predictions. The model enables clinicians to distinguish children requiring intervention from those with adequate growth potential despite current short stature. Clinicians can use these predicted trajectories as objective references for growth hormone therapy decisions and treatment monitoring.^8,10 When observed growth deviates from predicted patterns during follow-up, they can identify needs for reassessment or therapeutic adjustments.^17–19 Shapley additive explanations analysis enhances clinical interpretation by quantifying the relative contributions of body composition parameters to individual predictions. This integration of predictions with transparent reasoning supports data-driven clinical decisions while maintaining interpretive flexibility for complex cases. Moreover, the noninvasive nature of BIA enables frequent monitoring without radiation exposure and population-level growth surveillance through school-based screening programs. This scalability facilitates early identification of at-risk populations and provides epidemiological data for public health interventions.

One major limitation of our study is the absence of handwrist radiography data. This hindered our ability to directly compare our ML-based approach with conventional methods for evaluating skeletal maturity. Had we had access to this data, we could have assessed the added value of our model compared to traditional approaches and created a more comprehensive framework for predicting growth. Similarly, the unavailability of comparable longitudinal datasets with comprehensive body composition measurements prevented external validation of our model. Additionally, another limitation is the observational nature of the study, as we were unable to control for variables that could affect growth outcomes. As a result, the links identified between body composition and upcoming height through explainable AI methods should be interpreted with caution as they may not directly reflect causal relations.

Our model was developed exclusively using data from South Korean children and adolescents. Population-specific growth patterns represent a fundamental consideration for model generalizability. The onset of puberty, body composition trajectories, and adult height potential exhibit substantial variations across racial and ethnic groups.^58,59 These biological differences indicate that predictive relationships between body composition and height trajectory identified in Korean children may not directly translate to other populations. Future research should prioritize acquiring diverse ethnic datasets to conduct comprehensive external validation. More importantly, integrating multiethnic growth data with existing Korean cohort data could enable the development of population-adjusted prediction frameworks.

In addition, our research concentrated on children and adolescents aged 7–16 years due to limited data availability, which may restrict the applicability of our model for younger children or in predicting adult height. The reliability of our model for these populations could be compromised due to insufficient data on early growth trends, development after age 16, and the possible impact of various factors on adult height. Moreover, although our model exhibits strong accuracy and interpretability, integrating ML models into clinical workflows poses significant challenges. To enhance adoption, it is crucial to provide user-friendly interfaces and ensure smooth integration with electronic health record systems. Additionally, healthcare professionals will need training and support to accurately interpret and utilize the model's insights in their clinical decision-making. Finally, while our model encompasses a wide range of anthropometric and body composition metrics, it does not account for other genetic, environmental, or socioeconomic factors that may also play a role.

Despite these challenges, our research offers significant insights into the potential application of ML for assessing pediatric growth. By examining a large longitudinal dataset and considering various factors, our model generates tailored growth curves that provide interpretable insights, enhancing clinical decision-making. Future studies should focus on gathering comprehensive data, such as handwrist radiography, and utilize rigorous experimental methodologies to validate and expand our results. This study lays the essential foundation for further investigation into the application of ML in pediatric growth assessment, highlighting the need for continued interdisciplinary collaboration to refine these techniques and integrate them into clinical practice settings.

Conclusion

This study created an ML model to predict future height in children and adolescents by using anthropometric and body composition parameters. The model was trained on a large, longitudinal dataset, enabling it to forecast height and generate personalized growth curves accurately. Explainable AI methods were implemented to improve the interpretability of the model. While there are some limitations to this approach, it makes a significant contribution to pediatric growth assessment. It could serve as a valuable tool in clinical settings, helping to identify growth disorders and optimize outcomes.