An interpretable machine learning model based on routine clinical indicators for early identification of obesity-related non-alcoholic fatty liver disease in children: a single-center retrospective study

Introduction

Childhood obesity remains highly prevalent worldwide and has become a major public health challenge, accompanied by a continuous increase in the incidence of obesity-related metabolic complications (1,2). Among these, non-alcoholic fatty liver disease (NAFLD) has become one of the most common chronic liver diseases in children (3). Without timely intervention, some affected children may progress to non-alcoholic steatohepatitis (NASH) and liver fibrosis during childhood, and may even develop cirrhosis or hepatocellular carcinoma in adulthood (4,5). Therefore, early identification and risk assessment of NAFLD in children with obesity are of substantial clinical importance for preventing severe hepatic and extrahepatic complications.

At present, liver biopsy remains the gold standard for evaluating NASH and the degree of fibrosis. However, its invasive nature, relatively high medical cost, and potential complications such as bleeding limit its routine use in pediatric practice (6,7). Abdominal ultrasonography is widely used as a first-line screening tool for NAFLD because it is non-invasive, convenient, and readily accessible (8,9). Nevertheless, ultrasound findings are highly operator-dependent and have limited sensitivity for detecting mild hepatic steatosis. In particular, in children with severe obesity, the thickened subcutaneous fat layer may attenuate the ultrasound signal, thereby further compromising diagnostic accuracy (10). Therefore, there is an urgent clinical need to develop a quantitative assessment tool based on routinely available non-invasive clinical and biochemical indicators to facilitate the early identification of ultrasound-defined NAFLD risk in children with obesity.

In recent years, machine learning (ML) has shown distinct advantages in handling complex, high-dimensional clinical data and capturing nonlinear relationships among variables (11-13). Compared with traditional statistical models, ML methods have demonstrated superior predictive performance in certain clinical prediction tasks (14,15). However, some high-performing algorithms, especially ensemble learning models, are often regarded as “black boxes” because of their complex internal structures, which limits their interpretability and acceptability in clinical decision-making (16). To address this limitation, Shapley Additive exPlanations (SHAP) has been widely applied for model interpretation. This method can quantify and visualize the contribution of each clinical feature to the prediction outcome at both the global and individual levels, thereby improving the clinical interpretability of complex ML models (17,18).

Accordingly, the present study aimed to compare the predictive performance of multiple ML algorithms using real-world clinical cohort data and to identify and construct the optimal non-invasive clinical prediction model for ultrasound-defined NAFLD risk in children with obesity. In addition, SHAP analysis was introduced to interpret the best-performing model, thereby clarifying the direction and relative contribution of key clinical and biochemical features to NAFLD risk. This study was intended to provide a clinically useful auxiliary screening approach with both good predictive performance and interpretability for the early identification and risk assessment of ultrasound-defined fatty liver in children with obesity. We present this article in accordance with the TRIPOD reporting checklist (available at https://tp.amegroups.com/article/view/10.21037/tp-2026-0371/rc).

Methods

Study population and inclusion/exclusion criteria

This was a single-center retrospective clinical study that included children with obesity who attended the Department of Pediatrics at Siyang Hospital between February 2021 and December 2025. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Siyang Hospital (approval No. KS2026002). As this was a retrospective observational study, the requirement for informed consent from patients and their guardians was waived by the ethics committee.

The inclusion criteria were as follows: (I) age <18 years; (II) obesity defined according to the Chinese diagnostic criteria for childhood and adolescent obesity, namely a body mass index (BMI) at or above the 95th percentile for age and sex; (III) receipt of standardized abdominal ultrasonography during outpatient or inpatient care for the assessment of hepatic steatosis; and (IV) availability of complete demographic data, baseline clinical characteristics, and laboratory biochemical indicators.

The exclusion criteria were as follows: (I) presence of other definite acute or chronic liver diseases, such as viral hepatitis (hepatitis B or C), autoimmune liver disease, Wilson’s disease, or drug-induced liver injury; (II) secondary obesity caused by endocrine disorders, such as hypothyroidism or hypercortisolism, or genetic/metabolic syndromes such as Prader-Willi syndrome; (III) continuous use within 3 months before enrollment of drugs known to be hepatotoxic or to markedly affect glucose and lipid metabolism, such as systemic glucocorticoids, insulin sensitizers, or specific weight-loss medications; and (IV) excessive missingness in key variables that could not be reliably imputed, or incomplete abdominal ultrasound reports.

After screening, a total of 439 children with obesity were finally included as the data source for model development and internal testing.

Collection of clinical features and diagnostic criteria

Baseline clinical data for all enrolled participants were collected through the electronic medical record system. Demographic and anthropometric variables included sex, age, BMI, waist circumference (WC), waist-to-height ratio (WHtR), systolic blood pressure (SBP), and diastolic blood pressure (DBP).

After an overnight fast of at least 8 h, venous blood samples were collected from all participants the following morning according to a standardized protocol, and routine biochemical measurements were performed in the hospital laboratory. The measured indicators included glycated hemoglobin (HbA1c), fasting blood glucose, fasting insulin, 2 h post-oral glucose tolerance test (OGTT) glucose, 2 h post-OGTT insulin, triglycerides (TGs), low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C), serum uric acid, aspartate aminotransferase (AST), and alanine aminotransferase (ALT).

Based on these measurements, the following metabolism-related derived indices were further calculated: triglyceride-glucose (TyG) index = ln [fasting TG (mg/dL) × fasting glucose (mg/dL)/2]; TyG-WC (triglyceride-glucose index × waist circumference) = TyG index × WC (cm); TyG-BMI (triglyceride-glucose index × body mass index) = TyG index × BMI (kg/m²); and homeostasis model assessment of insulin resistance (HOMA-IR) = [fasting glucose (mmol/L) × fasting insulin (µIU/mL)]/22.5.

The study outcome was defined as ultrasound-defined hepatic steatosis (hereafter referred to as ultrasound-defined NAFLD risk) determined by abdominal ultrasonography. All abdominal ultrasound examinations were performed by experienced pediatric sonographers. According to relevant clinical consensus statements on pediatric NAFLD, the ultrasonographic diagnosis of hepatic steatosis consistent with NAFLD required the presence of at least two of the following typical imaging features: diffusely increased echogenicity of the near-field liver parenchyma that was stronger than that of the right renal cortex; blurred intrahepatic vascular or ductal structures; and gradual attenuation of far-field liver echogenicity. Given its non-invasive nature, convenience, and high clinical accessibility, abdominal ultrasonography was used in this study as the basis for defining hepatic steatosis.

Feature selection and data preprocessing

Considering that some baseline variables were missing in the real-world clinical dataset, and to minimize the selection bias associated with directly deleting incomplete cases, the following preprocessing was performed. First, continuous values more than three standard deviations (SDs) from the mean were set to missing. Missing values were then imputed using multivariate imputation by chained equations (the mice package in R, random seed 123), and a single completed dataset was used for all subsequent analyses; the overall proportion of missing data was low. The completed dataset was then divided into a training set and a testing set in a 7:3 ratio by stratified sampling (createDataPartition in R, random seed 123) to ensure a similar distribution of outcomes in both sets. Subsequently, feature selection and model training were performed only in the training set, whereas the test set was used solely for independent performance evaluation. To avoid data leakage, feature selection, and hyperparameter tuning were all restricted to the training workflow.

To identify robust predictive features associated with NAFLD in children with obesity, two complementary feature-selection methods were applied. First, least absolute shrinkage and selection operator (LASSO) regression was used, and the optimal penalty parameter (λ) was determined by 10-fold cross-validation to identify candidate variables with non-zero regression coefficients while reducing the risk of model overfitting. Next, the Boruta algorithm based on random forest was employed to identify outcome-related features by comparing the importance of original variables with that of shadow features.

To improve the stability of feature selection and the generalizability of the model, the intersection of features selected by LASSO regression and the Boruta algorithm was used as the final feature set. Ultimately, seven clinical features were included for subsequent model development: sex, BMI, WC, WHtR, serum uric acid, TyG-WC, and TyG-BMI.

Model development and evaluation

During model development, the study cohort was randomly split into a training set and atest set at a ratio of 7:3. Stratified sampling was used to ensure a consistent proportion of the outcome variable between the two sets.

Based on the seven key clinical features identified during feature selection, six commonly used ML classification models were developed and compared in the training set: logistic regression, decision tree, random forest, extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and support vector machine (SVM).

To optimize model performance and reduce overfitting, grid search combined with 5-fold cross-validation was used for hyperparameter tuning of each model to determine the optimal hyperparameter combination. The complete hyperparameter search grids for each model are provided [as follows: decision tree (max_depth 3/5/10/None, min_samples_split 5/10/20, max_features ’sqrt’/None, ccp_alpha 0/0.01/0.1); random forest (n_estimators 50–450, max_features 2–3); XGBoost (learning_rate 0.01/0.1/0.2, max_depth 3/5/10, n_estimators 50/100/200, subsample 0.6/0.8/1.0); LightGBM (learning_rate 0.1/0.2, num_leaves 31/50/100, n_estimators 100, subsample 0.6/0.8/1.0, colsample_bytree 0.6/0.8/1.0); and support vector machine (C 0.1, kernel linear/rbf, gamma scale/auto/0.01, degree 2/3/4). Logistic regression was used without additional tuning]. All class-based performance metrics were derived at a fixed probability threshold of 0.50, and no separate risk-stratification cut-offs were defined. The trained model objects and analysis code are available from the corresponding author on reasonable request. The predictive performance of each model was then evaluated in the test set. To characterize the variance of performance estimates and the stability of model ranking beyond the single split, we additionally performed repeated stratified cross-validation (5 folds × 20 repeats) on the full cohort, summarizing each metric as mean ± SD across folds (Table S1). The evaluation metrics included the area under the curve (AUC), accuracy, precision, sensitivity, specificity, F1 score, Cohen’s kappa, Youden’s J index, and negative predictive value (NPV).

Based on the overall comparison of these performance metrics, LightGBM was ultimately selected as the optimal prediction model for subsequent interpretability analysis.

Statistical analysis

Baseline descriptive statistics, between-group comparisons, and feature selection procedures (LASSO regression and Boruta algorithm) were performed in R software (version 4.3.1), primarily using the glmnet package (version 4.1-9) and the Boruta package (version 8.0.0). ML model construction, hyperparameter tuning, model performance evaluation, and SHAP-based interpretability analysis were conducted in Python (version 3.9.12), primarily using scikit-learn (version 1.2.2), lightgbm (version 3.3.5), xgboost (version 1.7.3), and shap (version 0.41.0).

For descriptive statistical analysis, the Shapiro-Wilk test was first used to assess the normality of continuous variables. Continuous variables with a normal distribution are presented as mean ± SD, and between-group comparisons were performed using the independent-samples t-test. Non-normally distributed continuous variables are presented as median and interquartile range (IQR), and comparisons were performed using the Mann-Whitney U test. Categorical variables are expressed as number (percentage), and comparisons were performed using the Chi-squared test or Fisher’s exact test, as appropriate. All statistical tests were two-sided, and a P value <0.05 was considered statistically significant. To assess multicollinearity among the selected predictors and the robustness of the SHAP-based importance ranking, we computed the Pearson correlation matrix and variance inflation factors, and additionally derived feature importance using two SHAP-independent approaches (permutation importance and LightGBM gain importance) together with an incremental analysis of nested feature sets using repeated cross-validation (Table S2).

Results

Baseline characteristics

A total of 439 children with obesity were included in this study, including 233 in the simple obesity group and 206 in the NAFLD group. The baseline characteristics are shown in Table 1. Compared with the simple obesity group, the NAFLD group showed statistically significant differences in sex distribution, BMI, WC, WHtR, SBP, HbA1c, TG, serum uric acid, and ALT (all P<0.05). In addition, the metabolic derived indices TyG index, TyG-WC, and TyG-BMI were also significantly higher in the NAFLD group (all P<0.05). No statistically significant differences were found between the two groups in age, DBP, fasting blood glucose, fasting insulin, 2 h post-OGTT glucose, 2 h post-OGTT insulin, LDL-C, HDL-C, AST, or HOMA-IR (all P>0.05).

Feature selection

To identify robust predictive features associated with NAFLD in children with obesity, 22 baseline clinical and biochemical variables were included in the feature-selection process. First, LASSO regression was applied. After determination of the optimal penalty parameter by 10-fold cross-validation, the LASSO model retained 13 candidate features with non-zero coefficients (Figure 1A-1D).

Figure 1 Feature selection using LASSO and Boruta. (A) Cross-validation curve for selection of the optimal penalty parameter (λ) in the LASSO model. (B) Coefficient profiles of candidate variables in the LASSO model as a function of log(λ). (C) Variable importance summary derived from the Boruta algorithm. (D) Variable importance history across Boruta iterations. ALT, alanine aminotransferase; AST, aspartate aminotransferase; BMI, body mass index; HbA1c, glycated hemoglobin; HOMA-IR, homeostatic model assessment of insulin resistance; LASSO, least absolute shrinkage and selection operator; LDL-C, low-density lipoprotein cholesterol; TG, triglyceride; TyG, triglyceride-glucose; WC, waist circumference; WHtR, waist-to-height ratio.

Meanwhile, variable importance was evaluated using the Boruta algorithm based on random forest. Through iterative comparison between the original variables and shadow features, the Boruta algorithm identified 11 important features.

To improve the stability of the selected features and the generalizability of subsequent models, the intersection of the LASSO- and Boruta-selected features was used as the final input set. Ultimately, seven core predictors were identified for subsequent ML model development: sex, BMI, WC, WHtR, serum uric acid, TyG-WC, and TyG-BMI.

Development and evaluation of ML models

Based on the seven core features identified through the dual feature-selection strategy, six ML classification models were developed in the training set, including logistic regression, decision tree, random forest, XGBoost, LightGBM, and SVM, and their predictive performance was evaluated in the test set. The ROC curves are shown in Figure 2. Random forest and LightGBM both achieved the highest AUC of 0.870 among all models. XGBoost and logistic regression also showed good discriminative ability, with AUCs of 0.862 and 0.853, respectively.

Figure 2 ROC curve comparison of six machine learning models in the test set. The gray dashed diagonal line indicates the reference line for random classification. AUC values with 95% CIs are provided in the legend. Random forest and LightGBM showed the highest discriminative performance (AUC =0.870), followed by XGBoost (AUC =0.862) and logistic regression (AUC =0.853). AUC, area under the curve; CI, confidence interval; LightGBM, light gradient boosting machine; ROC, receiver operating characteristic; SVM, support vector machine; XGBoost, extreme gradient boosting.

To further compare the overall performance of the models, multiple indicators, including accuracy, precision, sensitivity, specificity, and F1 score, were assessed (Table 2). Although LightGBM and random forest had the same AUC, LightGBM performed better across most of the other evaluation metrics. Specifically, LightGBM achieved the best performance in terms of accuracy, precision, sensitivity, F1 score, Cohen’s kappa, Youden’s J index, and NPV, while its specificity was tied with that of XGBoost (both 0.855). In contrast, the decision tree and SVM models showed relatively weaker overall performance. In repeated stratified cross-validation, the top models were closely clustered and statistically indistinguishable (mean AUC: XGBoost 0.828±0.038, random forest 0.822±0.040, LightGBM 0.807±0.040), consistent with the overlapping 95% confidence intervals on the test set (Figure 2). Because discrimination among the top models was comparable, LightGBM was selected on the test set for its best overall performance across the full set of evaluation metrics together with the best calibration (lowest Brier score, 0.146), rather than a uniquely higher AUC.

Model interpretability analysis

To assess the interpretability of the optimal LightGBM model, SHAP was applied at both the global and individual levels (Figures 3-5). The global feature-importance ranking and summary plot showed that WHtR contributed the most to the model output, followed by BMI, TyG-WC, serum uric acid, TyG-BMI, sex, and WC. In addition, higher values of WHtR, BMI, TyG-WC, serum uric acid, TyG-BMI, and WC were generally associated with higher SHAP values, indicating that increases in these features were related to an increased predicted risk of NAFLD.

Figure 3 SHAP-based interpretation of the LightGBM model. (A) SHAP waterfall plot for a representative high-risk case. (B) SHAP waterfall plot for a representative low-risk case. (C) Global SHAP feature importance plot of the LightGBM model, with features ranked by mean absolute SHAP values. BMI, body mass index; LightGBM, light gradient boosting machine; SHAP, SHapley Additive exPlanations; TyG, triglyceride-glucose; WC, waist circumference; WHtR, waist-to-height ratio.

Figure 4 SHAP summary plot of feature effects in the LightGBM model. Features are ranked by mean absolute SHAP value, including WHtR, BMI, TyG-WC, uric acid, TyG-BMI, gender, and WC. Each dot represents an individual sample and is colored according to the feature value. BMI, body mass index; LightGBM, light gradient boosting machine; SHAP, SHapley Additive exPlanations; TyG, triglyceride-glucose; WC, waist circumference; WHtR, waist-to-height ratio.

Figure 5 SHAP dependence plots of key feature effects in the LightGBM model. Each dot represents an individual sample. Red curves indicate fitted trends, and orange dashed lines denote the zero SHAP reference. BMI, body mass index; LightGBM, light gradient boosting machine; SHAP, SHapley Additive exPlanations; TyG, triglyceride-glucose; WC, waist circumference; WHtR, waist-to-height ratio.

The SHAP dependence plots further showed that BMI, serum uric acid, TyG-WC, TyG-BMI, and WC were overall positively associated with the model output. In contrast, the relationship between WHtR and the model output was clearly nonlinear: the corresponding SHAP values changed relatively slowly at low-to-moderate values, but increased markedly at higher values. The sex variable showed a relatively distinct stratified distribution pattern. Strong collinearity was limited to two mathematically related pairs (BMI-TyG-BMI, r=0.92; WC-TyG-WC, r=0.90), whereas WHtR, serum uric acid and sex were nearly independent (WHtR variance inflation factor =1.03). WHtR remained the highest-ranked predictor under both permutation importance and gain importance, reproducing the SHAP-based ranking. In the incremental analysis, adding WHtR to BMI increased the mean AUC from 0.678 to 0.758, whereas the metabolic-derived indices provided only a negligible increment beyond the anthropometric set (0.802 to 0.806) (Table S2).

At the individual level, the waterfall plots showed that the direction and magnitude of contribution of different clinical features varied across individual samples. For high-risk samples, BMI, WHtR, and serum uric acid were the main factors driving the predicted probability upward, while TyG-BMI, sex, TyG-WC, and WC also showed positive contributions to some extent. For low-risk samples, these features contributed in the opposite direction.

Model calibration and clinical utility

To evaluate clinical utility, calibration-curve analysis and decision-curve analysis were further performed in the test set (Figure 6A,6B). The calibration curves showed that LightGBM and random forest both had the lowest Brier score of 0.146 among all models, indicating the best calibration performance. The Brier scores for logistic regression, decision tree, XGBoost, and SVM were 0.158, 0.162, 0.170, and 0.181, respectively. Decision-curve analysis showed that, across a broad range of threshold probabilities, the net benefit of all models was generally higher than that of the “treat none” strategy, with LightGBM, random forest, and logistic regression demonstrating relatively higher overall net benefit. Taken together, these findings suggest that LightGBM has promising potential for clinical application.

Figure 6 Calibration and clinical utility analysis of six machine learning models in the test set. (A) Calibration curves of six machine learning models in the test set, with Brier scores and 95% CIs shown in the legend. (B) Decision curve analysis of six machine learning models in the test set, showing net benefit across threshold probabilities. CI, confidence interval; LightGBM, light gradient boosting machine; SVM, support vector machine; XGBoost, extreme gradient boosting.

Discussion

With the continuing global rise in childhood obesity and ongoing lifestyle changes, the burden of pediatric NAFLD is increasing and has become an important public health issue in pediatrics (19,20). However, there is still a lack of early screening tools that are simultaneously non-invasive, convenient, and cost-effective in clinical practice, making early identification and intervention challenging in children with obesity (21,22). Against this background, the present study developed and validated an ML-based prediction model for the non-invasive early identification of NAFLD in children with obesity using routinely available clinical indicators. By combining LASSO regression and the Boruta algorithm for feature selection, seven core predictors were ultimately identified: sex, BMI, WC, WHtR, serum uric acid, TyG-WC, and TyG-BMI. After systematic comparison of six mainstream ML algorithms, the LightGBM model showed relatively the best overall performance across multiple metrics, with comparable discrimination among the top models in internal testing (AUC =0.870), along with good calibration and potential clinical net benefit. Furthermore, by incorporating the SHAP interpretability framework, this study provided both global- and individual-level explanations of the optimal model. The results showed that WHtR was the feature contributing most to model prediction and exhibited a nonlinear association with model output.

Among the core predictors identified in this study, WHtR contributed the most to NAFLD risk in children with obesity. This finding suggests that, compared with the traditional indicator BMI, measures reflecting central fat accumulation may have greater discriminative value for identifying NAFLD risk in children. Previous studies have shown that visceral adipose tissue is highly metabolically active and may promote intrahepatic lipid accumulation by increasing the flux of free fatty acids to the liver, thereby further inducing oxidative stress and inflammatory responses (23). In addition, the SHAP dependence plot revealed a marked nonlinear relationship between WHtR and the model output: its effect on risk prediction remained relatively stable within the low-to-moderate range, but increased substantially once a higher range was reached. This finding further supports the key role of central obesity in the development and progression of pediatric NAFLD and suggests that WHtR may have potential threshold-warning value for identifying high-risk individuals. The leading contribution of WHtR was corroborated by two SHAP-independent importance measures and by its near-independence from the other predictors, and is therefore unlikely to be a collinearity-driven allocation artifact; accordingly, statements on the incremental value of the collinear metabolic-derived indices have been tempered.

In addition to WHtR, TyG-WC, TyG-BMI, and serum uric acid were also identified as important predictors in this study, further indicating that disorders of glucose-lipid metabolism and central obesity may jointly contribute to the development of pediatric NAFLD. As composite indicators derived from TGs and fasting glucose, the TyG index and its derivatives are considered sensitive markers of insulin resistance and may capture early metabolic abnormalities more effectively than single metabolic indicators (24,25). From a mechanistic perspective, insulin resistance may increase hepatic lipogenesis and be accompanied by impaired fatty acid oxidation and disordered lipid export, thereby aggravating abnormal lipid deposition in liver tissue. Meanwhile, previous studies have shown that hyperuricemia is closely associated with oxidative stress, endoplasmic reticulum stress, and inflammatory responses, and may interact with abdominal obesity and insulin resistance to promote the occurrence and progression of NAFLD (26-28).

Given that traditional logistic regression models mainly rely on linear assumptions and may not adequately characterize the complex high-dimensional nonlinear relationships involved in metabolic diseases, the present study further combined high-performance ML algorithms with interpretability analysis to improve predictive performance and clinical applicability. The results showed that the LightGBM model achieved the best overall performance across multiple metrics, with comparable discrimination among the top models. At the same time, the clinical utility of ML models depends not only on predictive accuracy but also on whether their outputs are sufficiently interpretable. Therefore, SHAP was further introduced to provide both global- and individual-level explanations for the LightGBM model. SHAP analysis not only revealed the relative importance of features but also demonstrated nonlinear relationships between key indicators such as WHtR and NAFLD risk. At the individual level, SHAP waterfall plots intuitively quantified the specific direction and magnitude of each variable’s contribution to the prediction for a given child. Therefore, the model developed in this study maintained good predictive performance while also showing substantial transparency and interpretability, thereby providing strong support for its potential clinical application.

Compared with previously reported pediatric ultrasound-defined NAFLD risk prediction tools that were often based on single indicators or traditional linear models, the present study incorporated metabolic derived indices such as TyG-WC and TyG-BMI and combined ensemble learning algorithms with the SHAP framework, thereby achieving a balance between predictive performance and model interpretability for risk assessment in children with obesity. Established non-invasive indices include the Hepatic Steatosis Index (HSI; based on the ALT/AST ratio, BMI, sex and diabetes; reported AUC ≈0.81), the Fatty Liver Index (FLI; based on TGs, BMI, gamma-glutamyltransferase and waist circumference; reported AUC ≈0.84), and the pediatric NAFLD fibrosis index (PNFI; based on age, waist circumference and TGs; reported AUC ≈0.85). These tools differ in target and applicability: HSI and FLI were developed in adults to predict ultrasound-defined steatosis, whereas the PNFI predicts liver fibrosis in children with established NAFLD—a different outcome from ours (29-31). Within our cohort, the HSI yielded an AUC of only 0.628 against the ultrasound-defined outcome (29), substantially lower than the proposed LightGBM model (test-set AUC 0.870); 66% of children already exceeded the HSI rule-in threshold (≥36), indicating that BMI-driven adult indices lose discrimination in a uniformly obese pediatric population. The FLI could not be computed because gamma-glutamyltransferase was not available in our dataset, and the PNFI was not benchmarked because it targets fibrosis rather than steatosis. Recent pediatric nomogram- and machine-learning-based models for childhood NAFLD/metabolic dysfunction-associated fatty liver disease (MAFLD) have also been proposed; compared with these single-index or linear approaches, the present study integrates central-adiposity and glucose-lipid indices within an interpretable ensemble framework. This strategy suggests that early identification of pediatric NAFLD should not rely solely on a single obesity-related indicator, but should also integrate information on glucose-lipid metabolic abnormalities and body fat distribution to improve the model’s ability to capture complex clinical phenotypes. Consistent with this, in our cohort the NAFLD group showed a higher median WHtR (0.64 vs. 0.59), higher serum uric acid (398.7 vs. 377.5 µmol/L) and higher TyG-WC (801.8 vs. 730.9) than the simple-obesity group, and the effect of WHtR on predicted risk was nonlinear, rising steeply only at higher values.

This study has several limitations. First, as a single-center retrospective study, it had a relatively limited sample size and could not completely avoid selection bias. Second, the outcome was defined on the basis of abdominal ultrasonography rather than higher-precision methods such as liver biopsy or magnetic resonance imaging, which may have led to some degree of misclassification. Furthermore, because ultrasound reports were obtained retrospectively as part of routine clinical care, the formal inter-rater reliability of the ultrasound reads could not be assessed, and the sonographers were not formally blinded to clinically available information; expectation bias therefore cannot be entirely excluded. Prospective designs with blinded, independently re-read images and reporting of inter-rater agreement are warranted in future studies. Third, only random splitting was used for internal testing, and no independent external validation was performed. Although repeated cross-validation confirmed that LightGBM consistently ranked among the top-performing models, the precise ordering of these models remains uncertain and requires confirmation in external, multicenter data. Therefore, the generalizability of the model still needs to be further evaluated in multicenter studies across different regions and populations. In addition, some included variables may have been strongly correlated with each other. In addition, gamma-glutamyltransferase was not available in our dataset, which precluded computation of the FLI for head-to-head comparison. Although this may improve predictive performance, it could affect the stability of model interpretation. In the future, multicenter prospective studies with larger sample sizes, more precise imaging or histopathological assessments, as well as external validation and model optimization, are needed to further improve the generalizability, stability, and clinical utility of the model.

Conclusions

Based on real-world clinical data from children with obesity, this study developed an ML model to predict the risk of ultrasound-defined NAFLD. After combined feature selection using LASSO regression and the Boruta algorithm, seven core predictors were ultimately included: sex, BMI, WC, WHtR, serum uric acid, TyG-WC, and TyG-BMI. Compared with the other models, LightGBM demonstrated superior overall predictive performance in internal testing, together with good calibration and potential clinical net benefit. SHAP interpretability analysis further indicated that WHtR, BMI, TyG-WC, serum uric acid, and TyG-BMI were important contributors to model prediction, with WHtR showing the greatest contribution, highlighting the importance of central obesity and glucose-lipid metabolic disturbances in identifying ultrasound-defined NAFLD risk in children with obesity. Overall, the LightGBM model developed in this study may serve as a non-invasive, convenient, and interpretable auxiliary tool to pre-screen children with obesity and inform referral for abdominal ultrasonography, supporting early identification and risk assessment of ultrasound-defined NAFLD; however, its clinical implementation still requires further confirmation through multicenter prospective studies and external validation.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tp.amegroups.com/article/view/10.21037/tp-2026-0371/rc

Data Sharing Statement: Available at https://tp.amegroups.com/article/view/10.21037/tp-2026-0371/dss

Peer Review File: Available at https://tp.amegroups.com/article/view/10.21037/tp-2026-0371/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tp.amegroups.com/article/view/10.21037/tp-2026-0371/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Siyang Hospital (approval No. KS2026002). As this was a retrospective observational study, the requirement for informed consent from patients and their guardians was waived by the ethics committee.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Zhang X, Liu J, Ni Y, et al. Global Prevalence of Overweight and Obesity in Children and Adolescents: A Systematic Review and Meta-Analysis. JAMA Pediatr 2024;178:800-13. [Crossref] [PubMed]
2. Park H, Choi JE, Jun S, et al. Metabolic complications of obesity in children and adolescents. Clin Exp Pediatr 2024;67:347-55. [Crossref] [PubMed]
3. Kalveram L, Baumann U, De Bruyne R, et al. Noninvasive scores are poorly predictive of histological fibrosis in paediatric fatty liver disease. J Pediatr Gastroenterol Nutr 2024;78:27-35. [Crossref] [PubMed]
4. Stroes AR, Vos M, Benninga MA, et al. Pediatric MASLD: current understanding and practical approach. Eur J Pediatr 2024;184:29. [Crossref] [PubMed]
5. Sergi CM. NAFLD (MASLD)/NASH (MASH): Does It Bother to Label at All? A Comprehensive Narrative Review. Int J Mol Sci 2024;25:8462.
6. Hunter AK, Lin HC. Review of Clinical Guidelines in the Diagnosis of Pediatric Nonalcoholic Fatty Liver Disease. Clin Liver Dis (Hoboken) 2021;18:40-4. [Crossref] [PubMed]
7. Chen BR, Pan CQ. Non-invasive assessment of fibrosis and steatosis in pediatric non-alcoholic fatty liver disease. Clin Res Hepatol Gastroenterol 2022;46:101755. [Crossref] [PubMed]
8. Ezenwuba BN, Hynes CM. Ultrasound screening of paediatric non-alcoholic fatty liver disease (NAFLD): A critical literature review. Radiography (Lond) 2024;30:1317-25. [Crossref] [PubMed]
9. Vos MB, Abrams SH, Barlow SE, et al. NASPGHAN Clinical Practice Guideline for the Diagnosis and Treatment of Nonalcoholic Fatty Liver Disease in Children: Recommendations from the Expert Committee on NAFLD (ECON) and the North American Society of Pediatric Gastroenterology, Hepatology and Nutrition (NASPGHAN). J Pediatr Gastroenterol Nutr 2017;64:319-34. [Crossref] [PubMed]
10. Kostrzeba E, Bik-Multanowski M, Brandt-Heunemann S, et al. Pediatric MASLD in Severe Obesity: Current Clinical Perspectives on Diagnosis and Treatment. J Clin Med 2025;15:137. [Crossref] [PubMed]
11. Runge JH, van Giessen J, Draijer LG, et al. Accuracy of controlled attenuation parameter compared with ultrasound for detecting hepatic steatosis in children with severe obesity. Eur Radiol 2021;31:1588-96. [Crossref] [PubMed]
12. Barrios JP, Tison GH. Advancing cardiovascular medicine with machine learning: Progress, potential, and perspective. Cell Rep Med 2022;3:100869. [Crossref] [PubMed]
13. Correa-Aguila R, Alonso-Pupo N, Hernández-Rodríguez EW. Multi-omics data integration approaches for precision oncology. Mol Omics 2022;18:469-79. [Crossref] [PubMed]
14. Liu W, Laranjo L, Klimis H, et al. Machine-learning versus traditional approaches for atherosclerotic cardiovascular risk prognostication in primary prevention cohorts: a systematic review and meta-analysis. Eur Heart J Qual Care Clin Outcomes 2023;9:310-22. [Crossref] [PubMed]
15. He Y, Liu N, Yang J, et al. Comparison of artificial intelligence and logistic regression models for mortality prediction in acute respiratory distress syndrome: a systematic review and meta-analysis. Intensive Care Med Exp 2025;13:23. [Crossref] [PubMed]
16. Laxar D, Eitenberger M, Maleczek M, et al. The influence of explainable vs non-explainable clinical decision support systems on rapid triage decisions: a mixed methods study. BMC Med 2023;21:359. [Crossref] [PubMed]
17. Allgaier J, Mulansky L, Draelos RL, et al. How does the model make predictions? A systematic literature review on the explainability power of machine learning in healthcare. Artif Intell Med 2023;143:102616.
18. Luo H, Xiang C, Zeng L, et al. SHAP based predictive modeling for 1 year all-cause readmission risk in elderly heart failure patients: feature selection and model interpretation. Sci Rep 2024;14:17728. [Crossref] [PubMed]
19. Lee EJ, Choi M, Ahn SB, et al. Prevalence of nonalcoholic fatty liver disease in pediatrics and adolescents: a systematic review and meta-analysis. World J Pediatr 2024;20:569-80. [Crossref] [PubMed]
20. Luo Y, Wang L, Yang X, et al. NAFLD in adolescents aged 10-19: Global Burden of Disease analysis from 1990 to 2021. Medicine (Baltimore) 2025;104:e45051. [Crossref] [PubMed]
21. Yodoshi T. Exploring non-invasive diagnostics and non-imaging approaches for pediatric metabolic dysfunction-associated steatotic liver disease. World J Gastroenterol 2024;30:5070-5. [Crossref] [PubMed]
22. Lin YC, Liao FM, Chao HCConsensus Statement on Metabolic Dysfunction-Associated Steatotic Liver Disease in Children and Adolescents From the Joint TASL-TSPGHAN Expert Committee, et al. JGH Open 2025;9:e70137. [Crossref] [PubMed]
23. Cortez APB, Fisberg M, de Morais MB. Intestinal permeability and small intestine bacterial overgrowth in excess weight adolescents. Pediatr Obes 2021;16:e12741. [Crossref] [PubMed]
24. Cabanillas-Lazo M, Quispe-Vicuña C, Pascual-Guevara M, et al. Diagnostic accuracy of triglyceride to glucose index and triglyceride/high-density lipoprotein index for insulin resistance among children and adolescents: A systematic review. PLoS One 2025;20:e0326179. [Crossref] [PubMed]
25. Trochimczyk K, Flisiak-Jackiewicz M, Bobrus-Chociej A, et al. Biochemical and Anthropometric Indices of Insulin Resistance in Obese and Overweight Children with Metabolic Dysfunction-Associated Fatty Liver Disease. Med Sci Monit 2024;30:e943375. [Crossref] [PubMed]
26. Nogueira JP, Cusi K. Role of Insulin Resistance in the Development of Nonalcoholic Fatty Liver Disease in People With Type 2 Diabetes: From Bench to Patient Care. Diabetes Spectr 2024;37:20-8. [Crossref] [PubMed]
27. Fan J, Wang D. Serum uric acid and nonalcoholic fatty liver disease. Front Endocrinol (Lausanne) 2024;15:1455132. [Crossref] [PubMed]
28. Zdanowicz K, Kopiczko N, Flisiak-Jackiewicz M, et al. Uric acid as a potential marker of cardiometabolic risk in children and adolescents with metabolic dysfunction associated steatotic liver disease. Clin Exp Hepatol 2024;10:188-93. [Crossref] [PubMed]
29. Lee JH, Kim D, Kim HJ, et al. Hepatic steatosis index: a simple screening tool reflecting nonalcoholic fatty liver disease. Dig Liver Dis 2010;42:503-8. [Crossref] [PubMed]
30. Bedogni G, Bellentani S, Miglioli L, et al. The Fatty Liver Index: a simple and accurate predictor of hepatic steatosis in the general population. BMC Gastroenterol 2006;6:33. [Crossref] [PubMed]
31. Nobili V, Alisi A, Vania A, et al. The pediatric NAFLD fibrosis index: a predictor of liver fibrosis in children with non-alcoholic fatty liver disease. BMC Med 2009;7:21. [Crossref] [PubMed]