Explainable machine learning model incorporating urinary heavy metals to predict nonalcoholic fatty liver disease

Explainable machine learning model incorporating urinary heavy metals to predict nonalcoholic fatty liver disease | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Explainable machine learning model incorporating urinary heavy metals to predict nonalcoholic fatty liver disease Xiaoqian Wang, Mei Xue, Hannah Chang, Bochun Wang, Wenquan Niu, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7286245/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Objectives: This study aimed to develop and validate an explainable machine learning (ML) model to predict NAFLD based on urinary heavy metals and phenotypic indices. Methods: Data were drawn from the NHANES 2017-2020. NAFLD was defined as a controlled attenuation parameter (CAP)≥274 dB/m. Urinary heavy metals were quantified by inductively coupled plasma mass spectrometry and normalized to urinary creatinine to account for dilution. Four ML algorithms (LightGBM, NNET, SVM, and XGBoost) were implemented. The dataset was split into training (60%) and validation (40%) sets. Results: Among 1,213 adults, 512 were classified with NAFLD and 701 as controls. XGBoost outperformed others, achieving superior performance (AUC=0.7983; Brier score=0.1804). Feature importance was assessed using SHapley Additive exPlanations (SHAP), identifying a minimal subset of 10 features that preserved model performance. The strongest predictors were: body roundness index, triglyceride, diabetes mellitus, sex, age, and urinary concentrations of cadmium, cesium, barium, lead, and tungsten. Both global and local SHAP interpretations validated these features' contributions. The optimized XGBoost model was deployed as a web application (https://wxqdepression.shinyapps.io/nafldapp/). Conclusions: XGBoost demonstrated superior performance in predicting NAFLD using a streamlined set of urinary heavy metals and phenotypic indicators. SHAP-based interpretability confirmed the relevance of this minimal feature set. Heavy metals Non-alcoholic fatty liver disease Machine learning SHapley Additive exPlanati Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 INTRODUCTION Non-alcoholic fatty liver disease (NAFLD) is the most common chronic liver disorder worldwide,[1] affecting an estimated 30.2% of the global population. Prevalence varies geographically, ranging from 16.1% in Australia to 34% in South America.[2] NAFLD encompasses a spectrum of liver pathology, from simple steatosis to non-alcoholic steatohepatitis (NASH),[3] , [4] which an subsequently progress to advanced fibrosis cirrhosis, and hepatocellular carcinoma (HCC). Notably, 15–25% of NASH cases advance to fibrosis or cirrhosis, with 3–5% of cirrhotic patients developing HCC annually.[5-7] Despite its growing burden, NAFLD often remains undiagnosed in clinical settings, largely due to limited awareness among primary healthcare providers and the lack of early, noninvasive screening tools. Alarmingly, recent studies suggest that up to 80% of advanced NAFLD cases are only detected after irreversible liver damage has occurred.[8] Given that NAFLD is largely preventable, there is an urgent need to better understand its etiology and to identify modifiable risk factors. Such insights are essential for developing effective prevention and early detection strategies at general population level. Environmental contaminants—particularly heavy metals—have emerged as important, yet underexplored, contributors to the development of NAFLD. At the cellular level, even low concentrations of heavy metals can induce oxidative stress, mitochondrial dysfunction and metabolic disturbances.[9] Exposure occurs through multiple routes, including contaminated water sources, inhalation of airborne particles, occupational settings, and consumption of crops grown in heavy metal-enriched soil.[10] Once accumulated, heavy metals may disrupt hepatic lipid metabolism via synergistic or antagonistic interactions,[11-13] thereby promoting NAFLD onset and progression. A review by Sadighara et al. highlighted consistent accusations between exposure to specific heavy metals—such as arsenic, cadmium, iron, lead, and mercury—and increased risk of fatty liver disease, while zinc and copper appeared to play protective roles in disease progression.[14] However, the role of heavy metals in NAFLD remains incompletely characterized. Most existing studies have assessed individual metals or limited combinations, often within geographically restricted populations.[15, 16] Moreover, traditional statistical methods (e.g., generalized linear model or weighted quantile sum regression) are limited in their ability to capture nonlinear relationships and high-dimensional interactions among multiple exposures. In contrast, machine learning (ML) techniques including random forest, support vector machine (SVM) and gradient boosting, offer powerful tools for modeling complex, multidimensional relationships in a data-driven manner. ML approaches can better account for nonlinearities, interactions, and heterogeneity, making them well suited for environmental health applications and personalized risk prediction. To date, however, no studies have applied ML models to investigated the associations between body burden of heavy metals and NAFLD risk in a nationally representative population. In this study, we aimed to develop a ML-based model to predict NAFLD using a comprehensive panel of 13 urinary heavy metals, alongside phenotypic indices including demographics, lifestyle factors, and comorbidities. Data were derived from a nationally representative U.S. adult cohort participating in the National Health and Nutrition Examination Survey (NHANES). To enhance interpretability, SHapley Additive exPlanations (SHAP) were applied to quantify the contribution of each feature to model predictions. This work addresses a critical knowledge gap by elucidating the relationship between environmental heavy metal exposure and NAFLD risk, while supporting future efforts in individual prevention and precision liver health strategies. METHODS Data source and study participants Data were obtained from NHANES, a continuous cross-sectional program launched in 1999 by the Centers for Disease Control and Prevention. NHANES is conducted in 2-year cycles and employs a complex, stratified, multistage probability sampling design to yield nationally representative estimates of the non-institutionalized U.S. civilian population. The NHANES protocol was approved by The National Center for Health Statistics (NCHS) Ethics Review Board. As this study involved secondary analysis of publicly available, de-identified data, it was exempted from institutional ethical review and informed consent requirements. This observational study adheres to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines.[17] For analysis, participants were drawn from the 2017-2020 NHANES cycles, during which urinary heavy metal concentrations were available (n=15,560). Exclusion criteria included: aged under 18 years (n=5,867), pregnancy (n=67), missing data on key metabolic indicators (n=5,782), missing urinary metal concentrations (n=2,593), and unavailable Controlled Attenuation Parameters (CAP) measurements (n=38). After applying these criteria, a total of 1,213 adult participants with complete data were included in the final analysis. NAFLD definition NAFLD was defined using CAP values obtained through vibration-controlled transient elastography (VCTE). CAP measurements were used to assess hepatic steatosis, with NAFLD defined as CAP ≥274 dB/m, based on established diagnostic thresholds.[18] Urinary metal and creatinine measurement Spot urine samples were collected at NHANES mobile examination centers, stored at −30°C and transported to the National Center for Environmental Health for centralized analysis. Totally, 13 urinary heavy metals were measured using Inductively Coupled Plasma Mass Spectrometry (ICP-MS) following a standardized dilution-based preparation protocol. The metals assayed included chromium (Cr), barium (Ba), cadmium (Cd), cobalt (Co), cesium (Cs), molybdenum (Mo), manganese (Mn), lead (Pb), antimony (Sb), tin (Sn), thallium (Tl), tungsten (W), and nickel (Ni). Urinary creatinine was quantified using an enzymatic method. To account for variations in urine concentration, heavy metal levels were normalized to urinary creatinine (expressed as μg/g creatinine). For metal concentrations below the limit of detection (LOD), NHANES protocol was followed by replacing values with LOD divided by the square root of two. Logarithms was applied to both urinary metal concentrations and creatinine levels to reduce skewness and stabilize variance. Phenotypic indexes Phenotypic indexes included demographic characteristics, lifestyle factors, anthropometric measures, laboratory findings, and comorbid conditions. Data were collected through standardized NHANES interviews and examinations. Demographic variables included age, sex, race/ethnicity (categorized as Mexican American, non-Hispanic Black, non-Hispanic White, and other). Socioeconomic indicators comprised education level (less than high school, high school graduate/general educational development [GED] or equivalent, and college or above), marital status (married, divorced or separated, or never married), and income-poverty ratio (PIR) status stratified as low, medium, and high. Lifestyle variables included smoking status (never, former, and current), alcohol consumption in the past year (yes and no), and physical activity (0, 1-2, 3-5, or >5 times/week). Anthropometric and Laboratory Measures were performed. The body roundness index (BRI), a validated marker of central obesity and visceral fat accumulation, was calculated using waist circumference and height, as defined by Thomas et al.[19] Compared to body mass index (BMI), BRI more accurately reflects visceral fat accumulation/central obesity and cardiometabolic risk.[20] Laboratory data included total cholesterol, triglyceride, urine creatinine, and liver stiffness measurements (LSM). LSM values were obtained via VCTE, with liver fibrosis defined as LSM ≥8 kPa.[21] Comorbidity data included diabetes mellitus, hypertension, and cardiovascular disease (CVD). Diabetes mellitus was defined by any of the following: self-reported physician diagnosis; use of insulin or glucose-lowering medications; glycated hemoglobin A1 ≥6.5%; fasting blood glucose ≥7 mmol/L; or 2-hour glucose ≥11.1 mmol/Lon oral glucose tolerance testing.[22] Hypertension was defined as: systolic blood pressure ≥130 mmHg, diastolic blood pressure ≥80 mmHg, or a self-reported history of physician-diagnosed hypertension.[23] CVD was defined as self-reported physician-diagnosed history of coronary heart disease, congestive heart failure, angina pectoris, or myocardial infarction. Statistical analysis An overview of the analytical pipeline is illustrated in Figure 1 . To develop and validate a ML-based prediction model for NAFLD, a three-step framework was employed: (1) data preprocessing, (2) model development and evaluation, (3) model interpretability and application deployment. Data preprocessing : The following steps were systematically applied to ensure data quality and model reliability: (i) Feature Exclusion: Variables with >30% missingness were excluded from the analysis; (ii) Outlier Treatment: Outliers were identified using the interquartile range (IQR) method (Q3+1.5×IQR) and flagged as missing; (iii) Multiple imputation: Five imputed datasets were generated using multiple imputation. The optimal dataset was selected based on the average area under the receiver operating characteristic curve (AUC) across 20 model-dataset combinations; (iv) Collinearity Reduction: Among highly correlated feature pairs (absolute Spearman’s ρ ≥0.8), one variable was removed based on lower clinical relevance or weaker association with NAFLD; (v) Multicollinearity Assessment: Features with a variance inflation factor (VIF) >5 were excluded to reduce multicollinearity; (vi) Feature Scaling: Continuous variables were standardized using Z-score transformation to ensure uniform feature scaling across models; (vii) Data Splitting: The dataset was randomly partitioned into a training set (n=728) and a validation set (n=485) at a 60:40 ratio, stratified to maintain balanced NAFLD prevalence across subsets. Model development and evaluation : Four ML algorithms were employed to construct predictive models for NAFLD: Light gradient boosting machine (LightGBM), neural network, support vector machine (SVM), and eXtreme Gradient Boosting (XGBoost). LightGBM and XGBoost are ensemble learning methods based on gradient boosting. LightGBM is optimized for memory efficiency and faster training by implementing histogram-based decision tree algorithms. XGBoost incorporates regularization techniques to prevent overfitting while maintaining high computational efficiency. Neural network are particularly well-suited for capturing complex, nonlinear relationships. They consist of multiple layers of interconnected artificial neurons inspired by biological neural networks. SVM is effective for small to medium-sized datasets and excels at learning complex non-linear decision boundaries through kernel functions. Model training was performed on the training set (n = 728), and performance was evaluated on the validation set (n = 485). Each algorithm was optimized using a grid search strategy across 30 hyperparameter combinations, with performance evaluated via five-fold cross-validation.[24, 25] The final tuned hyperparameters for all four models are summarized in Supplementary Table S1-S4. Model performance was assessed using a comprehensive set of 31 evaluation metrics, including: (1) Discrimination metrics: Accuracy, AUC, Balanced Accuracy, Binary Brier Score, Precision, Sensitivity, Specificity, F-beta Score, Matthews Correlation Coefficient (MCC), and Area under the Precision-Recall Curve (PR AUC). (2) Calibration and error metrics: Binary Brier Score, Classification Error, Logarithmic Loss, and Multiclass Brier Score. (3) Confusion matrix-derived metrics: True Positives, True Negatives, False Positives, False Negatives, Positive Predictive Value (PPV), Negative Predictive Value (NPV), False Positive Rate (FPR), False Negative Rate (FNR), False Discovery Rate (FDR), False Omission Rate (FOR), True Positive Rate (TPR), True Negative Rate (TNR). (4) Diagnostic utility metrics: Diagnostic Odds Ratio. (5) Multiclass performance (where applicable): Multiclass AUC using One-vs-One (Pairwise and Uniform Average), One-vs-Rest (Pairwise and Uniform Average), and Multiclass Extension. This exhaustive evaluation ensured robust benchmarking and reliable selection of the top-performing predictive model. Model interpretability : To enhance the transparency and clinical relevance of the machine learning models, we applied SHAP for model interpretation. SHAP is a game theory-based framework that quantifies the contribution of each input feature to individual predictions by computing Shapley values—a robust measure derived from cooperative game theory. Two key visualization tools were employed: (1) Waterfall plots: Display the cumulative impact of each feature on the prediction of a single participant, highlighting how feature values push the prediction above or below the baseline risk. (2) Force plots: Decompose an individual prediction into risk-increasing and risk-decreasing components, allowing for intuitive interpretation of personalized NAFLD risk profiles. These visualizations facilitate clinical insight into model decisions and support personalized risk communication. Application deployment : To enhance clinical utility, we developed an interactive web-based application using R Shiny framework. The deployed tool integrates the top 10 predictive features—comprising 5 phenotypic covariates and 5 urinary heavy metals. User can input individual-level data through a secure interface, which is processed on a remote server. The application provides: (1) Personalized NAFLD risk probability estimates based on the optimized XGBoost model; (2) Participant-specific SHAP waterfall plots that visualize the relative contribution of each feature to the predicted risk. This platform allows users and clinicians to explore how individual factors contribute to NAFLD risk, supporting personalized prevention strategies. Feature comparison : Descriptive statistics were used to compare features between NAFLD and non-NAFLD participants. Continuous variables were expressed as mean (SD) or median (IQR) based on normality. Categorical variables were summarized as counts and percentages (%). Independent t tests, Mann-Whitney U tests or χ 2 tests were used for between-group comparisons, as appropriate. A two-sided p <0.001 was considered statistically significant to account for multiple comparisons and reduce false positives. Software and statistical environment : All analyses were conducted in R (version 4.4.3) under RStudio Desktop (2023.12.1 Build 402, Ocean Storm Release). ML was performed using the mlr3proba and mlr3 packages (v3). SHAP analyses and visualizations were conducted using the kernelshap and shapviz packages. Data wrangling and visualization leveraged tidyverse and ggplot2 frameworks. RESULTS Baseline characteristics Baseline characteristics of the study participants stratified by NAFLD status are summarized in Table 1 . A total of 1,213 U.S. adults (616 men and 597 women) from the NHANES 2017−2020 cycle were included in the analysis, comprising 512 individuals with NAFLD and 701 non-NAFLD controls. Compared to participants without NAFLD, those with NAFLD were generally older, more likely to be male, of Mexican American ethnicity, and married. They exhibited higher prevalence of physical inactivity, elevated BMI, and comorbidities, such as diabetes mellitus, and hypertension. Model evaluation and selection All included features—comprising 13 urinary heavy medals and 16 phenotypic covariates— had missingness rates below 30%. The proportion of outliers ranged from 0.08% to 13.85%, and was treated as additional missing data. Missing values were addressed using multiple imputation, generating five complete datasets. Continuous variables were standardized across all datasets using Z-score transformation to ensure consistent feature scaling. Each imputed dataset was randomly divided into training and validation sets in a 60:40 ratio, maintaining balanced NAFLD prevalence. Four ML models—LightGBM, neural network, SVM, and XGBoost—were trained on the training sets and evaluated on the validation sets of each imputed dataset. Model performance was evaluated using 31 metrics, and average values across the five validation sets are summarized in Table 2 . Among the four models, XGBoost consistently demonstrated superior performance, achieving the highest area under the curve (AUC, 0.7983), F-beta score (0.6742), and sensitivity (0.6765), along with the lowest binary Brier score (0.1804). The ROC curves and AUC distributions further supported XGBoost’s superior discriminatory capacity ( Figure 2 ). For downstream interpretability and visualization, the second imputed dataset—with a median AUC across all five—was selected as the representative dataset. The ROC curves, precision-recall curves (PRCs), threshold-sensitivity curves, precision-threshold curves for each ML model in this dataset’s validation group are shown in Figure S1 . Corresponding confusion matrices are presented in Figure S8, with additional plots for the other four datasets provided in Supplementary Figures S2-S5 . Based on comprehensive and consistent evaluation across multiple metrics and datasets, XGBoost was selected as the optimal model for subsequent interpretation and application deployment. Feature selection The variable physical exercise was excluded from analysis due to a missingness rate exceeding 30%. Feature selection was subsequently performed using the top-performing XGBoost model on the second imputed dataset, with additional validation across the remaining four datasets. Feature correlation and collinearity were assessed using Spearman’s correlation coefficients ( Figures S6 – S7 ) and variance inflation factor (VIF) analysis (Table S5). A Spearman’s |ρ| ≥0.8 was considered indicative of high collinearity. A notable finding was the strong correlation between urinary creatinine and urinary manganese (|ρ|=0.8). Given the established role of manganese in metabolic syndrome,[26] and to retain biologically relevant exposures, urinary creatinine was excluded from the final feature set. All other variables demonstrated VIF values <5, indicating a low risk of multicollinearity. Thus, all features except physical exercise and urinary creatinine were retained for further analysis. To identify a parsimonious yet high-performing feature subset, SHAP values were calculated for all remaining features. Features were ranked in descending order of SHAP importance and cumulatively added to the model. Four key metrics—accuracy, area under the ROC curve (AUROC), F-beta score, and precision—were used to monitor changes in model performance. As shown in Figure S9 , performance improvements plateaued after the inclusion of the top ten features. This minimal feature subset, comprising five phenotypic covariates and five urinary heavy metals, was retained for downstream interpretability analyses and application deployment. Model interpretation The final XGBoost model, incorporating the top ten predictive features, was interpreted using SHAP to provide both global and local insights. Global Interpretability: At the global level, feature importance was ranked based on mean absolute SHAP values ( Figure 3A) . BRI emerged as the most influential predictor of NAFLD, followed by triglyceride, diabetes mellitus, sex, age, Cd, Cs, Ba, Pb, and W. A SHAP bee swarm plot illustrated the distribution and directionality of each feature’s contribution across all participants ( Figure 3B ). Positive SHAP values indicate that a feature increases the predicted probability of NAFLD, while negative values indicate a protective effect. For example, higher levels of BRI and triglyceride, presence of diabetes mellitus, and male sex were consistently associated with an increased risk of NAFLD. In contrast, lower levels of certain metals (e.g., Cd, Pb) exhibited negative SHAP values, suggesting potential inverse associations in this population. SHAP univariate dependence plots ( Figure 3C ) further revealed that age, BRI, Ba, Cs, W, and TG were positively correlated with SHAP values, while Cd and Pb showed negative trends. Local Interpretability: To illustrate the model’s individualized prediction process, local SHAP interpretation was conducted using waterfall and force plots for a representative participant from the validation set ( Figure 4 ). These plots decompose the prediction into additive contributions of each feature, shifting the base value (i.e., the model's average output) toward the final predicted probability. In this case, the participant was classified as having NAFLD, with a predicted probability of 0.761. The standardized feature values were: W: -2.43, triglyceride: 0.366, BRI: -0.79, Cd: -0.38, age: 0.567, Pb: 1.51, diabetes mellitus: 0, sex: 1, Ba: -0.646, and Cs: -0.737. In this individual, W and triglyceride had the strongest positive SHAP values, contributing most to the elevated NAFLD risk. Interestingly, despite a negative BRI and absence of diabetes, the overall risk remained high due to the influence of heavy metal exposures. The waterfall plot ranked the relative contribution of each feature and visualized the cumulative impact on the final prediction. These findings highlight the utility of SHAP-based interpretation for personalized risk stratification, offering a transparent and intuitive means to understand how both environmental exposures and clinical features jointly contribute to NAFLD prediction. This framework also supports the future development of precision nutrition and environmental risk mitigation strategies. Model stability To assess model robustness, five-fold cross-validation was performed across the second imputed dataset. SHAP value distributions remained highly consistent across the five independent subsets, as illustrated in Figure S 10 . This reproducibility underscores the stability and reliability of the identified key predictors, mitigating potential bias stemming from a single data partition and supporting the generalizability of model findings. Application deployment To enhance clinical applicability and facilitate personalized risk stratification, the final XGBoost model—trained on the second imputed dataset—was deployed as an interactive web-based application ( Figure 5 ). This tool allows users to input real-world values for the nine retained key features, including five urinary heavy metals and four phenotypic variables. The model instantly generates: An individualized NAFLD risk estimate, and A participant-specific SHAP waterfall plot, which visually explains how each input feature influences the predicted risk, either positively or negatively. This user-friendly interface provides a transparent and interpretable prediction tool that can support clinical decision-making, environmental exposure counseling, and precision nutrition strategies. The application is freely available at: https://wxqdepression.shinyapps.io/nafldapp/. DISCUSSION Using nationally representative data from NHANES 2017–2020 and a comprehensive evaluation of four machine learning models, we identified XGBoost as the most effective algorithm for predicting NAFLD among U.S. adults. Importantly, we derived a parsimonious set of ten key features—including five urinary heavy metals (cadmium, cesium, barium, lead, and tungsten) and five phenotypic indicators (age, sex, body roundness index, diabetes mellitus, and triglyceride levels)—that achieved predictive performance comparable to that of the full feature set. Global and local SHAP interpretability analyses confirmed the predictive robustness and transparency of this minimal feature set within the XGBoost model, bolstering model explainability and clinical trust. To support real-world implementation, we developed and deployed an interactive web-based application for individualized NAFLD risk assessment, integrating personalized input and SHAP-based visual explanations. To our knowledge, this is the first study to systematically evaluate the association between urinary heavy metals and NAFLD risk using machine learning integrated with SHAP-based explainability. These findings offer novel insights into environmental-metabolic interactions and provide a foundation for precision prevention strategies targeting NAFLD. The involvement of heavy metals in the pathophysiology of NAFLD is biologically plausible and increasingly supported by experimental and epidemiologic evidence.[27-29] Heavy metals exert pathogenic effects through multiple interconnected mechanisms, including oxidative stress-mediated hepatocellular injury, inflammatory pathway activation, disruption of lipid metabolism, and epigenetic reprogramming of hepatic gene expression.[30] [31] [32] [33] Emerging studies demonstrate that metals such as cadmium (Cd), lead (Pb), and arsenic (As) induce the generation of reactive oxygen species (ROS) in hepatocytes, initiating a cascade of damage characterized by lipid peroxidation, loss of membrane integrity, dysfunction of the mitochondrial electron transport chain, and depletion of intracellular antioxidants.[34, 35] Concurrently, these metals impair lipid homeostasis and insulin signaling. For instance, Cd downregulates PPAR-γ, a transcription factor essential for fatty acid oxidation, while upregulating SREBP-1c, a driver of de novo lipogenesis—together promoting hepatic triglyceride accumulation. Cd also interferes with insulin receptor substrate-1 (IRS-1) phosphorylation, impairing glucose uptake and enhancing hepatic gluconeogenesis.[36, 37] Additionally, heavy metals modulate adipokine signaling. Chronic Pb exposure, for example, has been associated with dysregulation of adiponectin and leptin, correlating with increased BMI and central adiposity.[38, 39] These converging pathways collectively underscore the contributory role of environmental heavy metal exposure in the development and progression of NAFLD, especially in susceptible populations. The association between heavy metals and NAFLD has been extensively investigated, though findings remain variable. A recent systematic review and meta-analysis by Pan et al. synthesized data on endocrine-disrupting chemicals and NAFLD, reporting a significant positive association between cadmium (Cd) exposure and NAFLD risk. However, this relationship showed substantial heterogeneity across geographic regions, biological sample types, and obesity status,[40] suggesting that population-specific exposure profiles and metabolic contexts may modulate the Cd–NAFLD association. Complementing these findings, a bioinformatics study by Zhang et al. validated the Cd–NAFLD link and further identified three cadmium toxicity targets—HCK, MYC, and DUSP6—as potential blood-based biomarkers for NAFLD diagnosis.[41] These molecular targets support cadmium’s role in hepatic oxidative stress and lipid metabolic dysregulation. The role of other metals remains controversial. For example, a randomized, double-blind, placebo-controlled trial by Moradi et al. found that chromium picolinate supplementation significantly reduced serum triglycerides, fasting insulin, HOMA-IR, and fetuin-A, indicating potential metabolic benefit. In contrast, a systematic review by Sadighara et al. found no consistent association between chromium exposure and the prevalence or progression of fatty liver disease.[14] These divergent findings may reflect heterogeneous genetic backgrounds, participant differences, limited statistical power, ignored interactions among heavy metals, or methodological inconsistencies across studies. To address these uncertainties and provide a more nuanced understanding, we leveraged the NHANES 2017–2020 dataset—a nationally representative U.S. cohort—to systematically evaluate the predictive utility of 13 urinary heavy metals for NAFLD. We implemented four state-of-the-art machine learning models and employed SHAP-based analyses to enhance model transparency and identify robust, interpretable predictors, thereby addressing critical knowledge gaps in environmental-metabolic interactions. It is noteworthy that among the four machine learning models evaluated, XGBoost consistently outperformed the others in predicting NAFLD, achieving an area under the ROC curve (AUC) approaching 80%. From both statistical and algorithmic standpoints, XGBoost demonstrated distinct advantages: it offers stronger regularization and more balanced parallelization compared to LightGBM; greater interpretability and computational efficiency than neural networks; and better adaptability to sparse features than SVM.[42-44] XGBoost is a highly optimized implementation of gradient boosting decision trees (GBDT), which builds additive models by sequentially training weak learners to minimize residual errors.[45] Unlike traditional GBDTs, XGBoost incorporates L1/L2 regularization, tree pruning, and advanced parallel computation, making it particularly robust to noise and effective in modeling nonlinear interactions—similar to neural networks, but with enhanced interpretability. Despite its strengths, XGBoost is not without limitations, including high computational cost, sensitivity to hyperparameters, and limited transparency when used in complex ensemble configurations. In this study, we mitigated these limitations by (1) curating a balanced dataset with 512 NAFLD cases and 701 controls; (2) optimizing hyperparameters through extensive grid search and cross-validation; and (3) enhancing interpretability via SHAP analysis, which allowed us to deconstruct model predictions into individual feature contributions. Beyond identifying XGBoost as the top-performing model, our multi-step feature selection approach distilled the full set of 29 variables to a parsimonious panel of ten key predictors—including five urinary heavy metals and five phenotypic covariates. Remarkably, this minimal feature set achieved comparable predictive performance to that of the full model, underscoring its efficiency, generalizability, and clinical applicability. Among the ten key predictors identified in this study, five were urinary heavy metals—cadmium (Cd), cesium (Cs), barium (Ba), lead (Pb), and tungsten (W)—each of which has been previously implicated in metabolic dysregulation and hepatic pathophysiology. Mechanistically, Cd disrupts mitochondrial function in hepatocytes, impairing oxidative phosphorylation and leading to ATP depletion.[46] Cs induces potassium loss, which exacerbates hepatic lipid peroxidation and contributes to the development of steatohepatitis.[47] Ba exposure elicits hepatic oxidative stress, characterized by elevated malondialdehyde (MDA) levels and depleted glutathione, indicative of cellular damage.[48] Pb accumulates in the liver and activates inflammatory pathways through NF-κB signaling, promoting the release of pro-inflammatory cytokines.[35] W, by acting as a competitive inhibitor of molybdenum cofactors, disrupts essential hepatic enzymes such as sulfite oxidase and xanthine oxidase, impairing detoxification and purine metabolism.[45] While most prior studies have focused on the impact of individual metals, the combined and interactive effects of metal mixtures on human metabolic health remain poorly understood. Our findings suggest that Cd, Cs, Ba, Pb, and W may act synergistically or additively to contribute to NAFLD pathogenesis. The identification of these metals as robust predictors in our ML model not only provides empirical support for their involvement in NAFLD but also points to novel environmental risk factors and potential therapeutic targets for prevention and early intervention. In addition to the five key heavy metals, our study underscores the exceptional predictive value of body roundness index (BRI)—a novel anthropometric metric designed to more accurately capture regional adiposity compared to traditional measures such as BMI. Among the ten key predictors identified, BRI emerged as the single strongest contributor to NAFLD risk in the XGBoost model. By design, BRI quantifies abdominal adiposity and ectopic fat deposition, both of which are strongly linked to hepatic steatosis through insulin resistance, lipotoxicity, and chronic low-grade inflammation.[19] These mechanisms underpin the role of visceral fat—as morphologically manifested by a rounded body shape—in promoting hepatic lipid accumulation and metabolic dysfunction. The robust predictive power of BRI observed in our model supports the pathophysiological relevance of central obesity in NAFLD development. Further research exploring the molecular and metabolic pathways linking BRI to NAFLD could deepen our mechanistic understanding and potentially establish BRI as a clinically useful marker for early identification and stratification of at-risk individuals. Several limitations of our study warrant consideration. First, the observational design of NHANES inherently limits our ability to establish causal relationships between urinary heavy metal exposure and NAFLD risk. Second, although machine learning models are well-suited for detecting complex, nonlinear interactions, the potential for residual confounding by unmeasured variables remains. ML-based inference should be complemented by mechanistic studies to confirm biological plausibility. Third, urinary heavy metals were measured at a single time point, which precludes evaluation of long-term or cumulative exposure—a key consideration in understanding environmental contributors to chronic liver disease. Fourth, because our analysis was conducted in a U.S.-based adult cohort, the generalizability of findings to populations with differing genetic, dietary, or sociocultural backgrounds remains uncertain. Future studies should pursue external validation in ethnically and geographically diverse cohorts. Fifth, while SHAP-based interpretability analyses enhance model transparency and highlight influential features, they do not substitute for toxicological or biological validation. Clinical studies are needed to determine threshold concentrations of heavy metals that confer meaningful NAFLD risk, and to elucidate the biological mechanisms underlying these associations. In this nationally representative cohort study, XGBoost demonstrated superior accuracy in predicting NAFLD compared to alternative machine learning models. Notably, a minimal feature set comprising five urinary heavy metals and five phenotypic indices achieved comparable predictive performance, with SHAP analyses providing robust interpretability and validation. To support clinical application, we developed a publicly accessible interactive web tool that generates individualized NAFLD risk profiles from routine clinical data. These findings highlight the promise of machine learning and urinary heavy metals as valuable tools in NAFLD risk stratification. Future efforts should focus on integrating interpretable ML models into clinical workflows to realize the full potential of AI-driven prevention strategies. Declarations Acknowledgements The authors would like to sincerely thank the NHANES participants and staff for contributing to the data collection and making the data available for public use. Funding information This work was supported by the Public Service Development and Reform Pilot Project of Beijing Medical Research Institute (W. Niu), the Capital’s Funds for Health Improvement and Research (Grant Number: 2024-2-1133). Competing Interests The authors have no relevant financial or non-financial interests to disclose. Author information Authors and Affiliations Capital Institute of Pediatrics, Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, China. Xiaoqian Wang Graduate School, Beijing University of Chinese Medicine, Beijing, China. Mei Xue China-Japan Friendship Hospital, Beijing, China. Mei Xue Precision, University of Michigan, Ann Arbor, Michigan, USA. Hannah Chang Northeast Forestry University, Harbin City, Heilongjiang Province, China. Bochun Wang Center for Evidence-Based Medicine, Capital Institute of Pediatrics, No.2 Yinghua East St., Chaoyang District, Beijing, 100029, China. Wenquan Niu Department of Medicine, University of Pittsburgh School of Medicine, 200 Meyran Avenue, Suite 200, Pittsburgh, PA, USA Chung-Chou H Chang Department of Biostatistics, University of Pittsburgh School of Public Health, 200 Meyran Avenue, Suite 200, Pittsburgh, PA, USA. Chung-Chou H Chang Precision Health Program, Department of Radiology, College of Human Medicine, Michigan State University, 766 Service Road, East Lansing, MI, USA. Xiaoqun Dong Contributions Xiaoqian Wang and Mei Xue designed the study and analyzed the data. Hannah Chang and Bochun Wang co-validated the analysis. The first draft was written by Xiaoqian Wang. Chung-Chou H Chang, Xiaoqun Dong, and Wenquan Niu were responsible for validating and revising the final content. All authors read and approved the final manuscript. Corresponding authors Correspondence to Chungchou H Chang, Xiaoqun Dong or Wenquan Niu. Ethics approval and consent to participate The data collected complied with the ethical guidelines established by the relevant institutions and/or national research councils, as well as the 1964 Declaration of Helsinki and its subsequent amendments or equivalent ethical standards. Informed consent was obtained from all individual participants included in the study. Consent for publication Not applicable. Availability of data and materials The National Health and Nutrition Examination Survey datasets are publicly available at the National Center for Health Statistics of the Center for Disease Control and Prevention (https:// www. cdc. gov/ nchs/ nhanes/ index. htm). Competing interests The authors declare no competing interests. References Riazi K, Azhari H, Charette JH, Underwood FE, King JA, Afshar EE, Swain MG, Congly SE, Kaplan GG, Shaheen AA: The prevalence and incidence of NAFLD worldwide: a systematic review and meta-analysis . Lancet Gastroenterol Hepatol 2022, 7 (9):851-861. Amini-Salehi E, Letafatkar N, Norouzi N, Joukar F, Habibi A, Javid M, Sattari N, Khorasani M, Farahmand A, Tavakoli S et al : Global Prevalence of Nonalcoholic Fatty Liver Disease: An Updated Review Meta-Analysis comprising a Population of 78 million from 38 Countries . Arch Med Res 2024, 55 (6):103043. Byrne CD, Targher G: NAFLD: a multisystem disease . J Hepatol 2015, 62 (1 Suppl):S47-64. Younossi ZM: Non-alcoholic fatty liver disease - A global public health perspective . J Hepatol 2019, 70 (3):531-544. McPherson S, Hardy T, Henderson E, Burt AD, Day CP, Anstee QM: Evidence of NAFLD progression from steatosis to fibrosing-steatohepatitis using paired biopsies: implications for prognosis and clinical management . J Hepatol 2015, 62 (5):1148-1155. Loomba R, Adams LA: The 20% Rule of NASH Progression: The Natural History of Advanced Fibrosis and Cirrhosis Caused by NASH . Hepatology 2019, 70 (6):1885-1888. Sheka AC, Adeyi O, Thompson J, Hameed B, Crawford PA, Ikramuddin S: Nonalcoholic Steatohepatitis: A Review . JAMA 2020, 323 (12):1175-1183. Zhang H, Zhou XD, Shapiro MD, Lip GYH, Tilg H, Valenti L, Somers VK, Byrne CD, Targher G, Yang W et al : Global burden of metabolic diseases, 1990-2021 . Metabolism 2024, 160 :155999. Kim JJ, Kim YS, Kumar V: Heavy metal toxicity: An update of chelating therapeutic strategies . J Trace Elem Med Biol 2019, 54 :226-231. Järup L: Hazards of heavy metal contamination . Br Med Bull 2003, 68 :167-182. Li W, Huang G, Tang N, Lu P, Jiang L, Lv J, Qin Y, Lin Y, Xu F, Lei D: Effects of heavy metal exposure on hypertension: A machine learning modeling approach . Chemosphere 2023, 337 :139435. Chang W, Li P: Copper and Diabetes: Current Research and Prospect . Mol Nutr Food Res 2023, 67 (23):e2300468. Wang Y, Wang Y, Li R, Ni B, Chen R, Huang Y, Cheng R, Li P, Li H, Peng Y et al : Low-grade systemic inflammation links heavy metal exposures to mortality: A multi-metal inflammatory index approach . Sci Total Environ 2024, 947 :174537. Sadighara P, Abedini AH, Irshad N, Ghazi-Khansari M, Esrafili A, Yousefi M: Association Between Non-alcoholic Fatty Liver Disease and Heavy Metal Exposure: a Systematic Review . Biol Trace Elem Res 2023, 201 (12):5607-5615. Zhang D, Wu S, Lan Y, Chen S, Wang Y, Sun Y, Lu Y, Liao W, Wang L: Essential metal mixtures exposure and NAFLD: A cohort-based case-control study in northern Chinese male adults . Chemosphere 2023, 339 :139598. Zhang D, Wu S, Lan Y, Chen S, Wang Y, Sun Y, Liao W, Wang L: Blood manganese and nonalcoholic fatty liver disease: A cohort-based case-control study . Chemosphere 2022, 287 (Pt 4):132316. Vandenbroucke JP, von Elm E, Altman DG, Gotzsche PC, Mulrow CD, Pocock SJ, Poole C, Schlesselman JJ, Egger M, Initiative S: Strengthening the Reporting of Observational Studies in Epidemiology (STROBE): explanation and elaboration . PLoS Med 2007, 4 (10):e297. Hsu C, Caussy C, Imajo K, Chen J, Singh S, Kaulback K, Le MD, Hooker J, Tu X, Bettencourt R et al : Magnetic Resonance vs Transient Elastography Analysis of Patients With Nonalcoholic Fatty Liver Disease: A Systematic Review and Pooled Analysis of Individual Participants . Clin Gastroenterol Hepatol 2019, 17 (4):630-637 e638. Thomas DM, Bredlau C, Bosy-Westphal A, Mueller M, Shen W, Gallagher D, Maeda Y, McDougall A, Peterson CM, Ravussin E et al : Relationships between body roundness with body fat and visceral adipose tissue emerging from a new geometrical model . Obesity (Silver Spring) 2013, 21 (11):2264-2271. Zhang X, Ma N, Lin Q, Chen K, Zheng F, Wu J, Dong X, Niu W: Body Roundness Index and All-Cause Mortality Among US Adults . JAMA Netw Open 2024, 7 (6):e2415051. de Lédinghen V, Wong VW, Vergniol J, Wong GL, Foucher J, Chu SH, Le Bail B, Choi PC, Chermak F, Yiu KK et al : Diagnosis of liver fibrosis and cirrhosis using liver stiffness measurement: comparison between M and XL probe of FibroScan® . J Hepatol 2012, 56 (4):833-839. ElSayed NA, Aleppo G, Aroda VR, Bannuru RR, Brown FM, Bruemmer D, Collins BS, Hilliard ME, Isaacs D, Johnson EL et al : 2. Classification and Diagnosis of Diabetes: Standards of Care in Diabetes-2023 . Diabetes Care 2023, 46 (Suppl 1):S19-s40. Whelton PK, Carey RM, Aronow WS, Casey DE, Jr., Collins KJ, Dennison Himmelfarb C, DePalma SM, Gidding S, Jamerson KA, Jones DW et al : 2017 ACC/AHA/AAPA/ABC/ACPM/AGS/APhA/ASH/ASPC/NMA/PCNA Guideline for the Prevention, Detection, Evaluation, and Management of High Blood Pressure in Adults: A Report of the American College of Cardiology/American Heart Association Task Force on Clinical Practice Guidelines . Hypertension 2018, 71 (6):e13-e115. Radzi SFM, Karim MKA, Saripan MI, Rahman MAA, Isa INC, Ibahim MJ: Hyperparameter Tuning and Pipeline Optimization via Grid Search Method and Tree-Based AutoML in Breast Cancer Prediction . J Pers Med 2021, 11 (10). Warkentin MT, Al-Sawaihey H, Lam S, Liu G, Diergaarde B, Yuan JM, Wilson DO, Atkar-Khattra S, Grant B, Brhane Y et al : Radiomics analysis to predict pulmonary nodule malignancy using machine learning approaches . Thorax 2024, 79 (4):307-315. Zhou J, Meng X, Deng L, Liu N: Non-linear associations between metabolic syndrome and four typical heavy metals: Data from NHANES 2011-2018 . Chemosphere 2022, 291 (Pt 2):132953. Canaday FT, Georas SN, Croft DP: Examining the impact of air pollution, climate change, and social determinants of health on asthma and environmental justice . Curr Opin Pulm Med 2024, 30 (3):276-280. Lavezzi AM, Ramos-Molina B: Environmental Exposure Science and Human Health . Int J Environ Res Public Health 2023, 20 (10). Hassanein EHM, Althagafy HS, Baraka MA, Amin H: Hepatoprotective effects of diosmin: a narrative review . Naunyn Schmiedebergs Arch Pharmacol 2025, 398 (1):279-295. Pan Z, Gong T, Liang P: Heavy Metal Exposure and Cardiovascular Disease . Circ Res 2024, 134 (9):1160-1178. Lamas GA, Bhatnagar A, Jones MR, Mann KK, Nasir K, Tellez-Plaza M, Ujueta F, Navas-Acien A: Contaminant Metals as Cardiovascular Risk Factors: A Scientific Statement From the American Heart Association . J Am Heart Assoc 2023, 12 (13):e029852. Could Your Mysterious Health Issues Be Heavy Metal Poisoning Symptoms? [https://www.ccfmed.com/blog/heavymetals] Koyama H, Kamogashira T, Yamasoba T: Heavy Metal Exposure: Molecular Pathways, Clinical Implications, and Protective Strategies . Antioxidants (Basel) 2024, 13 (1). Karri V, Schuhmacher M, Kumar V: Heavy metals (Pb, Cd, As and MeHg) as risk factors for cognitive dysfunction: A general review of metal mixture mechanism in brain . Environ Toxicol Pharmacol 2016, 48 :203-213. Thitame SN, Aher AA, Mopagar V: The Detoxifying Effects of Spirulina in Promoting Liver Health and Heavy Metal Elimination - A Review . J Pharm Bioallied Sci 2025, 17 (Suppl 1):S43-S45. Chen H, Tan H, Wan J, Zeng Y, Wang J, Wang H, Lu X: PPAR-gamma signaling in nonalcoholic fatty liver disease: Pathogenesis and therapeutic targets . Pharmacol Ther 2023, 245 :108391. Alruhaimi RS, Hassanein EHM, Alotaibi MF, Alzoghaibi MA, Abd El-Ghafar OAM, Mohammad MK, Alnasser SM, Mahmoud AM: Diallyl Disulfide Mitigates Cadmium Hepatotoxicity by Attenuating Oxidative Stress and TLR-4/NF-kappaB Signaling and Upregulating PPARgamma . Front Biosci (Landmark Ed) 2024, 29 (10):369. Mahmood S, Tanvir EM, Komarova T, Islam MN, Khatun M, Hossain MF, Ng JC, Whitfield KM, Hossain MS, Khalil MI et al : Relationships between growth indicators, liver and kidney function markers, and blood concentrations of essential and potentially toxic elements in environmentally exposed young children . Int J Hyg Environ Health 2023, 253 :114237. Little BB, Reilly R, Walsh B, Vu GT: Cadmium Is Associated with Type 2 Diabetes in a Superfund Site Lead Smelter Community in Dallas, Texas . Int J Environ Res Public Health 2020, 17 (12). Pan K, Xu J, Xu Y, Wang C, Yu J: The association between endocrine disrupting chemicals and nonalcoholic fatty liver disease: A systematic review and meta-analysis . Pharmacol Res 2024, 205 :107251. Zhang L, Wang R, Xue Q, Wang Y, Xu J, Wang C, Fang X, Gao S, Zhang H, Guo L: Bioinformatic Analysis for Exploring Target Genes and Molecular Mechanisms of Cadmium-Induced Nonalcoholic Fatty Liver Disease and Targeted Drug Prediction . J Appl Toxicol 2025, 45 (5):858-865. Deo RC: Machine Learning in Medicine . Circulation 2015, 132 (20):1920-1930. Zhang J, Ma X, Zhang J, Sun D, Zhou X, Mi C, Wen H: Insights into geospatial heterogeneity of landslide susceptibility based on the SHAP-XGBoost model . J Environ Manage 2023, 332 :117357. Dou B, Zhu Z, Merkurjev E, Ke L, Chen L, Jiang J, Zhu Y, Liu J, Zhang B, Wei GW: Machine Learning Methods for Small Data Challenges in Molecular Science . Chem Rev 2023, 123 (13):8736-8780. Christie IN, Theparambil SM, Braga A, Doronin M, Hosford PS, Brazhe A, Mascarenhas A, Nizari S, Hadjihambi A, Wells JA et al : Astrocytes produce nitric oxide via nitrite reduction in mitochondria to regulate cerebral blood flow during brain hypoxia . Cell Rep 2023, 42 (12):113514. Sun J, Chen Y, Wang T, Ali W, Ma Y, Yuan Y, Gu J, Bian J, Liu Z, Zou H: Cadmium promotes nonalcoholic fatty liver disease by inhibiting intercellular mitochondrial transfer . Cell Mol Biol Lett 2023, 28 (1):87. Racine R, Grandcolas L, Blanchardon E, Gourmelon P, Veyssiere G, Souidi M: Hepatic cholesterol metabolism following a chronic ingestion of cesium-137 starting at fetal stage in rats . J Radiat Res 2010, 51 (1):37-45. Elwej A, Chaabane M, Ghorbel I, Chelly S, Boudawara T, Zeghal N: Effects of barium graded doses on redox status, membrane bound ATPases and histomorphological aspect of the liver in adult rats . Toxicol Mech Methods 2017, 27 (9):677-686. Tables Table 1. Baseline characteristics of study participants Characteristics Normal controls (N=701) NAFLD patients (N=512) P-values Age (years) <0.001 18-39 282 (40.23) 107 (20.90) 40-59 195 (27.82) 188 (36.72) ≥60 224 (31.95) 217 (42.38) Men 334 (47.65) 282 (55.08) 0.011 Race/ethnicity <0.001 Mexican American 64 (9.13) 108 (21.09) Non-Hispanic Black 213 (30.39) 99 (19.34) Non-Hispanic White 228 (32.52) 172 (33.59) Other 196 (27.96) 133 (25.98) Marital status <0.001 Married 372 (56.53) 327 (65.14) Separated 145 (22.04) 109 (21.71) Never married 141 (21.43) 66 (13.15) Education level 0.094 Less than high school 117 (17.78) 111 (22.07) High school grad/GED 165 (25.08) 134 (26.64) College or above 376 (57.14) 258 (51.29) Income 0.345 Low 113 (18.77) 81 (18.28) Intermediate 243 (40.37) 198 (44.70) High 246 (40.46) 164 (37.02) Cigarette smoking 0.231 No smoking 393 (56.06) 298 (58.20) Former smoking 169 (24.11) 132 (25.78) Current smoking 139 (19.83) 82 (16.02) Alcohol drinking 0.979 No drinking 610 (91.45) 452 (91.50) Ever drinking 57 (8.55) 42 (8.50) Physical activity (times/week) 5 30 (6.22) 6 (1.75) BRI <0.001 0-20% 221 (31.53) 22 (4.30) 20-40% 186 (26.53) 57 (11.13) 40-60% 127 (18.12) 115 (22.46) 60-80% 91 (12.98) 152 (29.69) 80-100% 76 (10.84) 166 (32.42) DM <0.001 Yes 76 (10.84) 180 (35.16) No 625 (89.16) 332 (64.84) Hypertension <0.001 Yes 339 (51.36) 335 (69.50) No 321 (48.64) 147 (30.50) CVD 0.061 Yes 41 (6.25) 46 (9.18) No 615 (93.75) 455 (90.82) LSM ( kPa) <0.001 ≥ 8 37 (5.28) 92 (17.97) < 8 664 (94.72) 420 (82.03) Urinary creatinine (g/L), mean (SD) 0.061 ± 0.70 0.10 ± 0.61 0.023 TG (mg/dL), mean (SD) 85.16 ± 50.84 130.89 ± 71.27 <0.001 TC (mg/dL), mean (SD) 104.29 ± 33.94 112.12 ± 36.64 <0.001 Abbreviations: BRI, body roundness index; CVD, cardiovascular disease; DM, diabetes mellitus; GED, general educational development; LSM, liver serum marker; NAFLD, nonalcoholic fatty liver disease; PIR, ratio of family income to poverty; TC, total cholesterol; TG, triglyceride. Data are expressed as number (percentage) unless otherwise indicated. * For Income, low refers to PIR within [0, 1], intermediate within (1, 3], and high >3. ** P values were calculated using independent t test or Mann-Whitney U test for continuous variables and using χ 2 test for categorical variables, if appropriate. Table 2. Average 31 metrics derived from four machine learning models in validation groups of five imputed datasets Performance metrics LightGBM Neural network SVM XGBoost Accuracy 0.7282 0.7039 0.7126 0.7249 Area Under the ROC Curve 0.7937 0.7685 0.7648 0.7983 Balanced Accuracy 0.7201 0.6942 0.6963 0.7183 Binary Brier Score 0.1830 0.2033 0.1956 0.1804 Classification Error 0.2718 0.2961 0.2874 0.2751 Diagnostic Odds Ratio 6.8399 5.3448 5.8305 6.6370 Fβ-Score 0.6742 0.6427 0.6348 0.6742 False Discovery Rate 0.3201 0.3473 0.3183 0.3281 False Negatives 67.6 74.8 82.8 66.0 False Negative Rate 0.3314 0.3667 0.4059 0.3235 False Omission Rate 0.2376 0.2605 0.2695 0.2360 False Positives 64.2 68.8 56.6 67.4 False Positive Rate 0.2285 0.2448 0.2014 0.2399 Logarithmic Loss 0.5406 0.6041 0.5812 0.5346 Multiclass AUC Type 1 Pairwise 0.7937 0.7685 0.7648 0.7983 Multiclass AUC Type 1 Unweighted 0.7937 0.7685 0.7648 0.7983 Multiclass AUC Type N Pairwise 0.7937 0.7685 0.7648 0.7983 Multiclass AUC Type N Unweighted 0.7937 0.7685 0.7648 0.7983 Multiclass AUC Macro-Averaged 0.7937 0.7685 0.7648 0.7983 Multiclass Brier Score 0.3660 0.4067 0.3912 0.3607 Matthews Correlation Coefficient 0.4412 0.3903 0.4023 0.4363 Negative Predictive Value 0.7624 0.7395 0.7305 0.7640 Positive Predictive Value 0.6799 0.6527 0.6817 0.6719 Precision-Recall AUC 0.6980 0.6903 0.6908 0.7011 Precision 0.6799 0.6527 0.6817 0.6719 Sensitivity 0.6686 0.6333 0.5941 0.6765 Specificity 0.7715 0.7552 0.7986 0.7601 True Negatives 216.8 212.2 224.4 213.6 True Negative Rate 0.7715 0.7552 0.7986 0.7601 True Positives 136.4 129.2 121.2 138.0 True Positive Rate 0.6686 0.6333 0.5941 0.6765 Abbreviations: AUC, area under the ROC; LightGBM, light gradient boosting machine; ROC, receiver operating characteristic; SVM, support vector machine; XGBoost, eXtreme gradient boosting. Additional Declarations No competing interests reported. Supplementary Files Supplementarymaterials.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7286245","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":499040952,"identity":"5fc0ba27-6a57-4a0d-a570-73f464956157","order_by":0,"name":"Xiaoqian Wang","email":"","orcid":"","institution":"Chinese Academy of Medical Sciences \u0026 Peking Union Medical College","correspondingAuthor":false,"prefix":"","firstName":"Xiaoqian","middleName":"","lastName":"Wang","suffix":""},{"id":499040953,"identity":"026d267c-66b7-418f-9d85-1a377d17c74c","order_by":1,"name":"Mei Xue","email":"","orcid":"","institution":"Beijing University of Chinese Medicine","correspondingAuthor":false,"prefix":"","firstName":"Mei","middleName":"","lastName":"Xue","suffix":""},{"id":499040954,"identity":"b7676f00-ffd8-4152-9750-cc2fd2f986bd","order_by":2,"name":"Hannah Chang","email":"","orcid":"","institution":"University of Michigan","correspondingAuthor":false,"prefix":"","firstName":"Hannah","middleName":"","lastName":"Chang","suffix":""},{"id":499040955,"identity":"7800425a-9127-4f3b-a4ef-15f8b3fe82f1","order_by":3,"name":"Bochun Wang","email":"","orcid":"","institution":"Northeast Forestry University","correspondingAuthor":false,"prefix":"","firstName":"Bochun","middleName":"","lastName":"Wang","suffix":""},{"id":499040956,"identity":"c7cf3e57-04e8-443f-8a97-f33cc078fe32","order_by":4,"name":"Wenquan Niu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA6ElEQVRIiWNgGAWjYDACZjS2HBt7+wHStBjz8ZxJIM3GxHkSDgZ4VZmz85hJfNxhndg/u4H5c0HNnfQ2CYYEhh8V23BqsWzmMZOceSY9ccadA2zSM449y22TbjzA2HPmNk4tBod5zG7zth1ObLiRwMbM23A4t03mQAIzYxsBLX+BWubfSGD+DNSSziaRYEBYCyNQy4YbCQzSQC0JRGhhK//Z25ZuvBHoMGmeY4cN24CBfBCvX84f3mzws81adh7IYTw1h+Xl29sPPvhRgVsLFIBihP8DnHuAkHoG1DQwCkbBKBgFowANAAAgGlcViUHdUAAAAABJRU5ErkJggg==","orcid":"","institution":"Capital Institute of Pediatrics","correspondingAuthor":true,"prefix":"","firstName":"Wenquan","middleName":"","lastName":"Niu","suffix":""},{"id":499040957,"identity":"f57e88c6-01ce-4c5e-b744-4cee9bb17cfc","order_by":5,"name":"Chung-Chou H. Chang","email":"","orcid":"","institution":"University of Pittsburgh School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Chung-Chou","middleName":"H.","lastName":"Chang","suffix":""},{"id":499040958,"identity":"7f2ac735-f4ae-49c1-81d1-ca9b928bacbf","order_by":6,"name":"Xiaoqun Dong","email":"","orcid":"","institution":"Michigan State University","correspondingAuthor":false,"prefix":"","firstName":"Xiaoqun","middleName":"","lastName":"Dong","suffix":""}],"badges":[],"createdAt":"2025-08-04 02:38:32","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7286245/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7286245/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":89264252,"identity":"ad71ba17-87cc-4c62-b128-5b81a506a4ca","added_by":"auto","created_at":"2025-08-18 07:41:06","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":185795,"visible":true,"origin":"","legend":"\u003cp\u003eThe entire analytic flowchart of this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAbbreviations:\u003c/strong\u003eAPP, application; CAP, Controlled Attenuation Parameters; LightGBM, light gradient boosting machine; NHANES, National Health and Nutrition Examination Survey; ROC, receiver operating characteristic curve; ROSC, return of spontaneous circulation; SHAP, SHapley Additive exPlanations; SVM, support vector machine; XGBoost, eXtreme gradient boosting.\u003c/p\u003e","description":"","filename":"Picture1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7286245/v1/d6aa0ae846db36581233a1f1.jpg"},{"id":89264511,"identity":"9698b656-9918-4556-a117-3a4a04e91224","added_by":"auto","created_at":"2025-08-18 07:49:06","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":300211,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves (panel A) and AUC distributions (panel B) of four machine learning models for predicting nonalcoholic fatty liver disease based on the average values of validation groups from five imputed datasets\u003c/p\u003e\n\u003cp\u003eAbbreviations: AUC, area under the receiver operating characteristic \u003cstrong\u003e(\u003c/strong\u003eROC) curve; CI, confidence interval; FPR, false positive rate; TPR, true positive rate; LightGBM, light gradient boosting machine; SVM, support vector machine; XGBoost, eXtreme gradient boosting.\u003c/p\u003e","description":"","filename":"Picture2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7286245/v1/164cfec52cf4009fb78e6833.jpg"},{"id":89264512,"identity":"d8990e09-4861-4d9e-a7ee-e7956f6ffa45","added_by":"auto","created_at":"2025-08-18 07:49:06","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":674887,"visible":true,"origin":"","legend":"\u003cp\u003eGlobal SHAP interpretation of ten key features for predicting nonalcoholic fatty liver disease in validation group of the second imputed dataset using filtered mean ranking (panel A), bee swarm (panel B), and individual dependency plots (panel C)\u003c/p\u003e\n\u003cp\u003eAbbreviations: Ba,barium; BRI, body roundness index; Cd,cadmium; Cs,cesium; DM, diabetes mellitus; Pb,lead; TG, triglyceride;W,tungsten.\u003c/p\u003e","description":"","filename":"Picture3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7286245/v1/2bad79a33ceeec167bb95f0e.jpg"},{"id":89264514,"identity":"4555f679-8fe3-4cbf-89ee-151c252fac11","added_by":"auto","created_at":"2025-08-18 07:49:06","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":264836,"visible":true,"origin":"","legend":"\u003cp\u003eLocal SHAP interpretation of ten key features for predicting nonalcoholic fatty liver disease in validation group of the second imputed dataset using waterfall (panel A) and force (panel B) plots\u003c/p\u003e\n\u003cp\u003eAbbreviations: Ba,barium; BRI, body roundness index; Cd,cadmium; Cs,cesium; DM, diabetes mellitus; Pb,lead; TG, triglyceride;W,tungsten.\u003c/p\u003e","description":"","filename":"Picture4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7286245/v1/5333a625b96f5ce80da9a4d0.jpg"},{"id":89264513,"identity":"2c050946-fae6-4aff-a6cd-c5c3b01adec6","added_by":"auto","created_at":"2025-08-18 07:49:06","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":70714,"visible":true,"origin":"","legend":"\u003cp\u003eDevelopment and deployment of a web application for predicting nonalcoholic fatty liver disease\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7286245/v1/be2f711b89a5338d88742a7f.jpg"},{"id":91252846,"identity":"1a10ecf2-b388-42f5-aa4d-34f82c13a2f8","added_by":"auto","created_at":"2025-09-13 17:46:42","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4698301,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7286245/v1/2795e12a-5009-46c0-b31f-68fe52d0e5a8.pdf"},{"id":89264255,"identity":"43e51d88-19f7-40fd-85f9-49cd570404d8","added_by":"auto","created_at":"2025-08-18 07:41:06","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":1401351,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-7286245/v1/5d01b7aabf5a0bfacc79ce5f.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Explainable machine learning model incorporating urinary heavy metals to predict nonalcoholic fatty liver disease","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eNon-alcoholic fatty liver disease (NAFLD) is the most common chronic liver disorder worldwide,[1] affecting an estimated 30.2% of the global population. Prevalence varies geographically, ranging from 16.1% in Australia to 34% in South America.[2] NAFLD encompasses a spectrum of liver pathology, from simple steatosis to non-alcoholic steatohepatitis (NASH),[3]\u003csup\u003e,\u003c/sup\u003e[4] which an subsequently progress to advanced fibrosis cirrhosis, and hepatocellular carcinoma (HCC). Notably, 15–25% of NASH cases advance to fibrosis or cirrhosis, with 3–5% of cirrhotic patients developing HCC annually.[5-7] Despite its growing burden, NAFLD often remains undiagnosed in clinical settings, largely due to limited awareness among primary healthcare providers and the lack of early, noninvasive screening tools. Alarmingly, recent studies suggest that up to 80% of advanced NAFLD cases are only detected after irreversible liver damage has occurred.[8] Given that NAFLD is largely preventable, there is an urgent need to better understand its etiology and to identify modifiable risk factors. Such insights are essential for developing effective prevention and early detection strategies at general population level.\u003c/p\u003e\n\u003cp\u003eEnvironmental contaminants—particularly heavy metals—have emerged as important, yet underexplored, contributors to the development of NAFLD. At the cellular level, even low concentrations of heavy metals can induce oxidative stress, mitochondrial dysfunction and metabolic disturbances.[9] Exposure occurs through multiple routes, including contaminated water sources, inhalation of airborne particles, occupational settings, and consumption of crops grown in heavy metal-enriched soil.[10] Once accumulated, heavy metals may disrupt hepatic lipid metabolism via synergistic or antagonistic interactions,[11-13] thereby promoting NAFLD onset and progression. A review by Sadighara et al. highlighted consistent accusations between exposure to specific heavy metals—such as arsenic, cadmium, iron, lead, and mercury—and increased risk of fatty liver disease, while zinc and copper appeared to play protective roles in disease progression.[14] However, the role of heavy metals in NAFLD remains incompletely characterized. Most existing studies have assessed individual metals or limited combinations, often within geographically restricted populations.[15, 16] Moreover, traditional statistical methods (e.g., generalized linear model or weighted quantile sum regression) are limited in their ability to capture nonlinear relationships and high-dimensional interactions among multiple exposures. In contrast, machine learning (ML) techniques including random forest, support vector machine (SVM) and gradient boosting, offer powerful tools for modeling complex, multidimensional relationships in a data-driven manner. ML approaches can better account for nonlinearities, interactions, and heterogeneity, making them well suited for environmental health applications and personalized risk prediction. To date, however, no studies have applied ML models to investigated the associations between body burden of heavy metals and NAFLD risk in a nationally representative population.\u003c/p\u003e\n\u003cp\u003eIn this study, we aimed to develop a ML-based model to predict NAFLD using a comprehensive panel of 13 urinary heavy metals, alongside phenotypic indices including demographics, lifestyle factors, and comorbidities. Data were derived from a nationally representative U.S. adult cohort participating in the National Health and Nutrition Examination Survey (NHANES). To enhance interpretability, SHapley Additive exPlanations (SHAP) were applied to quantify the contribution of each feature to model predictions.\u0026nbsp;This work addresses a critical knowledge gap by elucidating the relationship between environmental heavy metal exposure and NAFLD risk, while supporting future efforts in individual prevention and precision liver health strategies.\u0026nbsp;\u003c/p\u003e"},{"header":"METHODS","content":"\u003cp\u003e\u003cstrong\u003eData source and study participants\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData were obtained from NHANES, a continuous cross-sectional program launched in 1999 by the Centers for Disease Control and Prevention. NHANES is conducted in 2-year cycles and employs a complex, stratified, multistage probability sampling design to yield nationally representative estimates of the non-institutionalized U.S. civilian population. The NHANES protocol was approved by The National Center for Health Statistics (NCHS) Ethics Review Board. As this study involved secondary analysis of publicly available, de-identified data, it was exempted from institutional ethical review and informed consent requirements. This observational study adheres to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines.[17]\u003c/p\u003e\n\u003cp\u003eFor analysis, participants were drawn from the 2017-2020 NHANES cycles, during which urinary heavy metal concentrations were available (n=15,560). Exclusion criteria included: aged under 18 years (n=5,867), pregnancy (n=67), missing data on key metabolic indicators (n=5,782), missing urinary metal concentrations (n=2,593), and unavailable Controlled Attenuation Parameters (CAP) measurements (n=38). After applying these criteria, a total of 1,213 adult participants with complete data were included in the final analysis.\u003c/p\u003e\n\u003cp skip=\"true\"\u003e\u003cstrong\u003eNAFLD definition\u003c/strong\u003e\u003c/p\u003e\n\u003cp skip=\"true\"\u003eNAFLD was defined using CAP values obtained through vibration-controlled transient elastography (VCTE). CAP measurements were used to assess hepatic steatosis, with NAFLD defined as CAP \u0026ge;274 dB/m, based on established diagnostic thresholds.[18]\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eUrinary metal and creatinine measurement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSpot urine samples were collected at NHANES mobile examination centers, stored at \u0026minus;30\u0026deg;C and transported to the National Center for Environmental Health for centralized analysis. Totally, 13 urinary heavy metals were measured using Inductively Coupled Plasma Mass Spectrometry (ICP-MS) following a standardized dilution-based preparation protocol. The metals assayed included chromium (Cr), barium (Ba), cadmium (Cd), cobalt (Co), cesium (Cs), molybdenum (Mo), manganese (Mn), lead (Pb), antimony (Sb), tin (Sn), thallium (Tl), tungsten (W), and nickel (Ni). Urinary creatinine was quantified using an enzymatic method. To account for variations in urine concentration, heavy metal levels were normalized to urinary creatinine (expressed as \u0026mu;g/g creatinine). For metal concentrations below the limit of detection (LOD), NHANES protocol was followed by replacing values with LOD divided by the square root of two. Logarithms was applied to both urinary metal concentrations and creatinine levels to reduce skewness and stabilize variance.\u0026nbsp;\u003c/p\u003e\n\u003cp skip=\"true\"\u003e\u003cstrong\u003ePhenotypic indexes\u003c/strong\u003e\u003c/p\u003e\n\u003cp skip=\"true\"\u003ePhenotypic indexes included demographic characteristics, lifestyle factors, anthropometric measures, laboratory findings, and comorbid conditions. Data were collected through standardized NHANES interviews and examinations. Demographic variables included age, sex, race/ethnicity (categorized as Mexican American, non-Hispanic Black, non-Hispanic White, and other). Socioeconomic indicators comprised education level (less than high school, high school graduate/general educational development [GED] or equivalent, and college or above), marital status (married, divorced or separated, or never married), and income-poverty ratio (PIR) status stratified as low, medium, and high. Lifestyle variables included smoking status (never, former, and current), alcohol consumption in the past year (yes and no), and physical activity (0, 1-2, 3-5, or \u0026gt;5 times/week). Anthropometric and Laboratory Measures were performed. The body roundness index (BRI), a validated marker of central obesity and visceral fat accumulation, was calculated using waist circumference and height, as defined by Thomas et al.[19] Compared to body mass index (BMI), BRI more accurately reflects visceral fat accumulation/central obesity and cardiometabolic risk.[20] Laboratory data included total cholesterol, triglyceride, urine creatinine, and liver stiffness measurements (LSM). LSM values were obtained via VCTE, with liver fibrosis defined as LSM \u0026ge;8 kPa.[21] Comorbidity data included diabetes mellitus, hypertension, and cardiovascular disease (CVD). Diabetes mellitus was defined by any of the following: self-reported physician diagnosis; use of insulin or glucose-lowering medications; glycated hemoglobin A1 \u0026ge;6.5%; fasting blood glucose \u0026ge;7 mmol/L; or 2-hour glucose \u0026ge;11.1 mmol/Lon oral glucose tolerance testing.[22] Hypertension was defined as: systolic blood pressure \u0026ge;130 mmHg, diastolic blood pressure \u0026ge;80 mmHg, or a self-reported history of physician-diagnosed hypertension.[23] CVD was defined as self-reported physician-diagnosed history of coronary heart disease, congestive heart failure, angina pectoris, or myocardial infarction.\u003c/p\u003e\n\u003cp skip=\"true\"\u003e\u003cstrong\u003eStatistical analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAn overview of the analytical pipeline is illustrated in \u003cstrong\u003eFigure 1\u003c/strong\u003e. To develop and validate a ML-based prediction model for NAFLD, a three-step framework was employed: (1) data preprocessing, (2) model development and evaluation, (3) model interpretability and application deployment.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eData preprocessing\u003c/em\u003e\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e The following steps were systematically applied to ensure data quality and model reliability: (i) Feature Exclusion: Variables with \u0026gt;30% missingness were excluded from the analysis; (ii) Outlier Treatment: Outliers were identified using the interquartile range (IQR) method (\u0026lt;Q1\u0026ndash;1.5\u0026times;IQR [interquartile range] or \u0026gt;Q3+1.5\u0026times;IQR) and flagged as missing; (iii) Multiple imputation: Five imputed datasets were generated using multiple imputation. The optimal dataset was selected based on the average area under the receiver operating characteristic curve (AUC) across 20 model-dataset combinations; (iv) Collinearity Reduction: Among highly correlated feature pairs (absolute Spearman\u0026rsquo;s \u0026rho; \u0026ge;0.8), one variable was removed based on lower clinical relevance or weaker association with NAFLD; (v) Multicollinearity Assessment: Features with a variance inflation factor (VIF) \u0026gt;5 were excluded to reduce multicollinearity; (vi) Feature Scaling: Continuous variables were standardized using Z-score transformation to ensure uniform feature scaling across models; (vii) Data Splitting: The dataset was randomly partitioned into a training set (n=728) and a validation set (n=485) at a 60:40 ratio, stratified to maintain balanced NAFLD prevalence across subsets.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eModel development and evaluation\u003c/em\u003e\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e Four ML algorithms were employed to construct predictive models for NAFLD: Light gradient boosting machine (LightGBM), neural network, support vector machine (SVM), and eXtreme Gradient Boosting (XGBoost). LightGBM and XGBoost are ensemble learning methods based on gradient boosting. LightGBM is optimized for memory efficiency and faster training by implementing histogram-based decision tree algorithms. XGBoost incorporates regularization techniques to prevent overfitting while maintaining high computational efficiency. Neural network are particularly well-suited for capturing complex, nonlinear relationships. They consist of multiple layers of interconnected artificial neurons inspired by biological neural networks. SVM is effective for small to medium-sized datasets and excels at learning complex non-linear decision boundaries through kernel functions.\u003c/p\u003e\n\u003cp\u003eModel training was performed on the training set (n = 728), and performance was evaluated on the validation set (n = 485). Each algorithm was optimized using a grid search strategy across 30 hyperparameter combinations, with performance evaluated via five-fold cross-validation.[24, 25] The final tuned hyperparameters for all four models are summarized in Supplementary Table S1-S4. Model performance was assessed using a comprehensive set of 31 evaluation metrics, including: (1) Discrimination metrics: Accuracy, AUC, Balanced Accuracy, Binary Brier Score, Precision, Sensitivity, Specificity, F-beta Score, Matthews Correlation Coefficient (MCC), and Area under the Precision-Recall Curve (PR AUC). (2) Calibration and error metrics: Binary Brier Score, Classification Error, Logarithmic Loss, and Multiclass Brier Score. (3) Confusion matrix-derived metrics: True Positives, True Negatives, False Positives, False Negatives, Positive Predictive Value (PPV), Negative Predictive Value (NPV), False Positive Rate (FPR), False Negative Rate (FNR), False Discovery Rate (FDR), False Omission Rate (FOR), True Positive Rate (TPR), True Negative Rate (TNR). (4) Diagnostic utility metrics: Diagnostic Odds Ratio. (5) Multiclass performance (where applicable): Multiclass AUC using One-vs-One (Pairwise and Uniform Average), One-vs-Rest (Pairwise and Uniform Average), and Multiclass Extension.\u0026nbsp;This exhaustive evaluation ensured robust benchmarking and reliable selection of the top-performing predictive model.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eModel interpretability\u003c/em\u003e\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e To enhance the transparency and clinical relevance of the machine learning models, we applied SHAP for model interpretation. SHAP is a game theory-based framework that quantifies the contribution of each input feature to individual predictions by computing Shapley values\u0026mdash;a robust measure derived from cooperative game theory. Two key visualization tools were employed: (1) Waterfall plots: Display the cumulative impact of each feature on the prediction of a single participant, highlighting how feature values push the prediction above or below the baseline risk. (2) Force plots: Decompose an individual prediction into risk-increasing and risk-decreasing components, allowing for intuitive interpretation of personalized NAFLD risk profiles. These visualizations facilitate clinical insight into model decisions and support personalized risk communication.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eApplication deployment\u003c/em\u003e\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e To enhance clinical utility, we developed an interactive web-based application using R Shiny framework. The deployed tool integrates the top 10 predictive features\u0026mdash;comprising 5 phenotypic covariates and 5 urinary heavy metals. User can input individual-level data through a secure interface, which is processed on a remote server. The application provides: (1) Personalized NAFLD risk probability estimates based on the optimized XGBoost model; (2) Participant-specific SHAP waterfall plots that visualize the relative contribution of each feature to the predicted risk. This platform allows users and clinicians to explore how individual factors contribute to NAFLD risk, supporting personalized prevention strategies.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eFeature comparison\u003c/em\u003e\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e Descriptive statistics were used to compare features between NAFLD and non-NAFLD participants. Continuous variables were expressed as mean (SD) or median (IQR) based on normality. Categorical variables were summarized as counts and percentages (%). Independent t tests, Mann-Whitney U tests or \u0026chi;\u003csup\u003e2\u003c/sup\u003e tests were used for between-group comparisons, as appropriate. A two-sided \u003cem\u003ep\u003c/em\u003e \u0026lt;0.001 was considered statistically significant to account for multiple comparisons and reduce false positives.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eSoftware and statistical environment\u003c/em\u003e\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e All analyses were conducted in R (version 4.4.3) under RStudio Desktop (2023.12.1 Build 402, Ocean Storm Release). ML was performed using the mlr3proba and mlr3 packages (v3). SHAP analyses and visualizations were conducted using the kernelshap and shapviz packages. Data wrangling and visualization leveraged tidyverse and ggplot2 frameworks.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cp skip=\"true\"\u003e\u003cstrong\u003eBaseline characteristics\u003c/strong\u003e\u003c/p\u003e\n\u003cp skip=\"true\"\u003eBaseline characteristics of the study participants stratified by NAFLD status are summarized in \u003cstrong\u003eTable 1\u003c/strong\u003e. A total of 1,213 U.S. adults (616 men and 597 women) from the NHANES 2017\u0026minus;2020 cycle were included in the analysis, comprising 512 individuals with NAFLD and 701 non-NAFLD controls. Compared to participants without NAFLD, those with NAFLD were generally older, more likely to be male, of Mexican American ethnicity, and married. They exhibited higher prevalence of physical inactivity, elevated BMI, and comorbidities, such as diabetes mellitus, and hypertension.\u003c/p\u003e\n\u003cp skip=\"true\"\u003e\u003cstrong\u003eModel evaluation and selection\u003c/strong\u003e\u003c/p\u003e\n\u003cp skip=\"true\"\u003eAll included features\u0026mdash;comprising 13 urinary heavy medals and 16 phenotypic covariates\u0026mdash; had missingness rates below 30%. The proportion of outliers ranged from 0.08% to 13.85%, and was treated as additional missing data. Missing values were addressed using multiple imputation, generating five complete datasets. Continuous variables were standardized across all datasets using Z-score transformation to ensure consistent feature scaling. Each imputed dataset was randomly divided into training and validation sets in a 60:40 ratio, maintaining balanced NAFLD prevalence.\u003c/p\u003e\n\u003cp skip=\"true\"\u003eFour ML models\u0026mdash;LightGBM, neural network, SVM, and XGBoost\u0026mdash;were trained on the training sets and evaluated on the validation sets of each imputed dataset. Model performance was evaluated using 31 metrics, and average values across the five validation sets are summarized in \u003cstrong\u003eTable 2\u003c/strong\u003e. Among the four models, XGBoost consistently demonstrated superior performance, achieving the highest area under the curve (AUC, 0.7983), F-beta score (0.6742), and sensitivity (0.6765), along with the lowest binary Brier score (0.1804). The ROC curves and AUC distributions further supported XGBoost\u0026rsquo;s superior discriminatory capacity (\u003cstrong\u003eFigure 2\u003c/strong\u003e). For downstream interpretability and visualization, the second imputed dataset\u0026mdash;with a median AUC across all five\u0026mdash;was selected as the representative dataset. The ROC curves, precision-recall curves (PRCs), threshold-sensitivity curves, precision-threshold curves for each ML model in this dataset\u0026rsquo;s validation group are shown in \u003cstrong\u003eFigure S1\u003c/strong\u003e. Corresponding confusion matrices are presented in \u003cstrong\u003eFigure S8,\u003c/strong\u003e with additional plots for the other four datasets provided in Supplementary \u003cstrong\u003eFigures S2-S5\u003c/strong\u003e. Based on comprehensive and consistent evaluation across multiple metrics and datasets, XGBoost was selected as the optimal model for subsequent interpretation and application deployment.\u003c/p\u003e\n\u003cp skip=\"true\"\u003e\u003cstrong\u003eFeature selection\u003c/strong\u003e\u003c/p\u003e\n\u003cp skip=\"true\"\u003eThe variable physical exercise was excluded from analysis due to a missingness rate exceeding 30%. Feature selection was subsequently performed using the top-performing XGBoost model on the second imputed dataset, with additional validation across the remaining four datasets. Feature correlation and collinearity were assessed using Spearman\u0026rsquo;s correlation coefficients (\u003cstrong\u003eFigures S6\u003c/strong\u003e\u003cstrong\u003e\u0026ndash;\u003c/strong\u003e\u003cstrong\u003eS7\u003c/strong\u003e) and variance inflation factor (VIF) analysis (Table S5). A Spearman\u0026rsquo;s |\u0026rho;| \u0026ge;0.8 was considered indicative of high collinearity. A notable finding was the strong correlation between urinary creatinine and urinary manganese (|\u0026rho;|=0.8). Given the established role of manganese in metabolic syndrome,[26] and to retain biologically relevant exposures, urinary creatinine was excluded from the final feature set. All other variables demonstrated VIF values \u0026lt;5, indicating a low risk of multicollinearity. Thus, all features except physical exercise and urinary creatinine were retained for further analysis.\u003c/p\u003e\n\u003cp skip=\"true\"\u003eTo identify a parsimonious yet high-performing feature subset, SHAP values were calculated for all remaining features. Features were ranked in descending order of SHAP importance and cumulatively added to the model. Four key metrics\u0026mdash;accuracy, area under the ROC curve (AUROC), F-beta score, and precision\u0026mdash;were used to monitor changes in model performance. As shown in \u003cstrong\u003eFigure S9\u003c/strong\u003e, performance improvements plateaued after the inclusion of the top ten features. This minimal feature subset, comprising five phenotypic covariates and five urinary heavy metals, was retained for downstream interpretability analyses and application deployment.\u003c/p\u003e\n\u003cp skip=\"true\"\u003e\u003cstrong\u003eModel interpretation\u003c/strong\u003e\u003c/p\u003e\n\u003cp skip=\"true\"\u003eThe final XGBoost model, incorporating the top ten predictive features, was interpreted using SHAP to provide both global and local insights. Global Interpretability: At the global level, feature importance was ranked based on mean absolute SHAP values (\u003cstrong\u003eFigure 3A)\u003c/strong\u003e. BRI emerged as the most influential predictor of NAFLD, followed by triglyceride, diabetes mellitus, sex, age, Cd, Cs, Ba, Pb, and W. A SHAP bee swarm plot illustrated the distribution and directionality of each feature\u0026rsquo;s contribution across all participants (\u003cstrong\u003eFigure 3B\u003c/strong\u003e).\u003cstrong\u003e \u003c/strong\u003ePositive SHAP values indicate that a feature increases the predicted probability of NAFLD, while negative values indicate a protective effect. For example, higher levels of BRI and triglyceride, presence of diabetes mellitus, and male sex were consistently associated with an increased risk of NAFLD. In contrast, lower levels of certain metals (e.g., Cd, Pb) exhibited negative SHAP values, suggesting potential inverse associations in this population. SHAP univariate dependence plots (\u003cstrong\u003eFigure 3C\u003c/strong\u003e) further revealed that age, BRI, Ba, Cs, W, and TG were positively correlated with SHAP values, while Cd and Pb showed negative trends.\u003c/p\u003e\n\u003cp skip=\"true\"\u003eLocal Interpretability: To illustrate the model\u0026rsquo;s individualized prediction process, local SHAP interpretation was conducted using waterfall and force plots for a representative participant from the validation set (\u003cstrong\u003eFigure 4\u003c/strong\u003e). These plots decompose the prediction into additive contributions of each feature, shifting the base value (i.e., the model\u0026apos;s average output) toward the final predicted probability. In this case, the participant was classified as having NAFLD, with a predicted probability of 0.761. The standardized feature values were: W: -2.43, triglyceride: 0.366, BRI: -0.79, Cd: -0.38, age: 0.567, Pb: 1.51, diabetes mellitus: 0, sex: 1, Ba: -0.646, and Cs: -0.737. In this individual, W and triglyceride had the strongest positive SHAP values, contributing most to the elevated NAFLD risk. Interestingly, despite a negative BRI and absence of diabetes, the overall risk remained high due to the influence of heavy metal exposures. The waterfall plot ranked the relative contribution of each feature and visualized the cumulative impact on the final prediction. These findings highlight the utility of SHAP-based interpretation for personalized risk stratification, offering a transparent and intuitive means to understand how both environmental exposures and clinical features jointly contribute to NAFLD prediction. This framework also supports the future development of precision nutrition and environmental risk mitigation strategies.\u003c/p\u003e\n\u003cp skip=\"true\"\u003e\u003cstrong\u003eModel stability\u003c/strong\u003e\u003c/p\u003e\n\u003cp skip=\"true\"\u003eTo assess model robustness, five-fold cross-validation was performed across the second imputed dataset. SHAP value distributions remained highly consistent across the five independent subsets, as illustrated in \u003cstrong\u003eFigure S\u003c/strong\u003e\u003cstrong\u003e10\u003c/strong\u003e. This reproducibility underscores the stability and reliability of the identified key predictors, mitigating potential bias stemming from a single data partition and supporting the generalizability of model findings.\u003c/p\u003e\n\u003cp skip=\"true\"\u003e\u003cstrong\u003eApplication deployment\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo enhance clinical applicability and facilitate personalized risk stratification, the final XGBoost model\u0026mdash;trained on the second imputed dataset\u0026mdash;was deployed as an interactive web-based application (\u003cstrong\u003eFigure 5\u003c/strong\u003e). This tool allows users to input real-world values for the nine retained key features, including five urinary heavy metals and four phenotypic variables. The model instantly generates: An individualized NAFLD risk estimate, and A participant-specific SHAP waterfall plot, which visually explains how each input feature influences the predicted risk, either positively or negatively. This user-friendly interface provides a transparent and interpretable prediction tool that can support clinical decision-making, environmental exposure counseling, and precision nutrition strategies. The application is freely available at: https://wxqdepression.shinyapps.io/nafldapp/.\u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eUsing nationally representative data from NHANES 2017–2020 and a comprehensive evaluation of four machine learning models, we identified XGBoost as the most effective algorithm for predicting NAFLD among U.S. adults. Importantly, we derived a parsimonious set of ten key features—including five urinary heavy metals (cadmium, cesium, barium, lead, and tungsten) and five phenotypic indicators (age, sex, body roundness index, diabetes mellitus, and triglyceride levels)—that achieved predictive performance comparable to that of the full feature set. Global and local SHAP interpretability analyses confirmed the predictive robustness and transparency of this minimal feature set within the XGBoost model, bolstering model explainability and clinical trust. To support real-world implementation, we developed and deployed an interactive web-based application for individualized NAFLD risk assessment, integrating personalized input and SHAP-based visual explanations. To our knowledge, this is the first study to systematically evaluate the association between urinary heavy metals and NAFLD risk using machine learning integrated with SHAP-based explainability. These findings offer novel insights into environmental-metabolic interactions and provide a foundation for precision prevention strategies targeting NAFLD.\u003c/p\u003e\n\u003cp\u003eThe involvement of heavy metals in the pathophysiology of NAFLD is biologically plausible and increasingly supported by experimental and epidemiologic evidence.[27-29] Heavy metals exert pathogenic effects through multiple interconnected mechanisms, including oxidative stress-mediated hepatocellular injury, inflammatory pathway activation, disruption of lipid metabolism, and epigenetic reprogramming of hepatic gene expression.[30] [31] [32] [33]\u0026nbsp;Emerging studies demonstrate that metals such as cadmium (Cd), lead (Pb), and arsenic (As) induce the generation of reactive oxygen species (ROS) in hepatocytes, initiating a cascade of damage characterized by lipid peroxidation, loss of membrane integrity, dysfunction of the mitochondrial electron transport chain, and depletion of intracellular antioxidants.[34, 35]\u0026nbsp;Concurrently, these metals impair lipid homeostasis and insulin signaling. For instance, Cd downregulates PPAR-γ, a transcription factor essential for fatty acid oxidation, while upregulating SREBP-1c, a driver of de novo lipogenesis—together promoting hepatic triglyceride accumulation. Cd also interferes with insulin receptor substrate-1 (IRS-1) phosphorylation, impairing glucose uptake and enhancing hepatic gluconeogenesis.[36, 37] Additionally, heavy metals modulate adipokine signaling. Chronic Pb exposure, for example, has been associated with dysregulation of adiponectin and leptin, correlating with increased BMI and central adiposity.[38, 39] These converging pathways collectively underscore the contributory role of environmental heavy metal exposure in the development and progression of NAFLD, especially in susceptible populations.\u003c/p\u003e\n\u003cp\u003eThe association between heavy metals and NAFLD has been extensively investigated, though findings remain variable. A recent systematic review and meta-analysis by Pan et al. synthesized data on endocrine-disrupting chemicals and NAFLD, reporting a significant positive association between cadmium (Cd) exposure and NAFLD risk. However, this relationship showed substantial heterogeneity across geographic regions, biological sample types, and obesity status,[40] suggesting that population-specific exposure profiles and metabolic contexts may modulate the Cd–NAFLD association. Complementing these findings, a bioinformatics study by Zhang et al. validated the Cd–NAFLD link and further identified three cadmium toxicity targets—HCK, MYC, and DUSP6—as potential blood-based biomarkers for NAFLD diagnosis.[41]\u0026nbsp;These molecular targets support cadmium’s role in hepatic oxidative stress and lipid metabolic dysregulation. The role of other metals remains controversial. For example, a randomized, double-blind, placebo-controlled trial by Moradi et al. found that chromium picolinate supplementation significantly reduced serum triglycerides, fasting insulin, HOMA-IR, and fetuin-A, indicating potential metabolic benefit. In contrast, a systematic review by Sadighara et al. found no consistent association between chromium exposure and the prevalence or progression of fatty liver disease.[14]\u0026nbsp;These divergent findings may reflect heterogeneous genetic backgrounds, participant differences, limited statistical power, ignored interactions among heavy metals, or methodological inconsistencies across studies. To address these uncertainties and provide a more nuanced understanding, we leveraged the NHANES 2017–2020 dataset—a nationally representative U.S. cohort—to systematically evaluate the predictive utility of 13 urinary heavy metals for NAFLD. We implemented four state-of-the-art machine learning models and employed SHAP-based analyses to enhance model transparency and identify robust, interpretable predictors, thereby addressing critical knowledge gaps in environmental-metabolic interactions.\u003c/p\u003e\n\u003cp\u003eIt is noteworthy that among the four machine learning models evaluated, XGBoost consistently outperformed the others in predicting NAFLD, achieving an area under the ROC curve (AUC) approaching 80%. From both statistical and algorithmic standpoints, XGBoost demonstrated distinct advantages: it offers stronger regularization and more balanced parallelization compared to LightGBM; greater interpretability and computational efficiency than neural networks; and better adaptability to sparse features than SVM.[42-44] XGBoost is a highly optimized implementation of gradient boosting decision trees (GBDT), which builds additive models by sequentially training weak learners to minimize residual errors.[45] Unlike traditional GBDTs, XGBoost incorporates L1/L2 regularization, tree pruning, and advanced parallel computation, making it particularly robust to noise and effective in modeling nonlinear interactions—similar to neural networks, but with enhanced interpretability. Despite its strengths, XGBoost is not without limitations, including high computational cost, sensitivity to hyperparameters, and limited transparency when used in complex ensemble configurations. In this study, we mitigated these limitations by (1) curating a balanced dataset with 512 NAFLD cases and 701 controls; (2) optimizing hyperparameters through extensive grid search and cross-validation; and (3) enhancing interpretability via SHAP analysis, which allowed us to deconstruct model predictions into individual feature contributions. Beyond identifying XGBoost as the top-performing model, our multi-step feature selection approach distilled the full set of 29 variables to a parsimonious panel of ten key predictors—including five urinary heavy metals and five phenotypic covariates. Remarkably, this minimal feature set achieved comparable predictive performance to that of the full model, underscoring its efficiency, generalizability, and clinical applicability.\u003c/p\u003e\n\u003cp\u003eAmong the ten key predictors identified in this study, five were urinary heavy metals—cadmium (Cd), cesium (Cs), barium (Ba), lead (Pb), and tungsten (W)—each of which has been previously implicated in metabolic dysregulation and hepatic pathophysiology. Mechanistically, Cd disrupts mitochondrial function in hepatocytes, impairing oxidative phosphorylation and leading to ATP depletion.[46] Cs induces potassium loss, which exacerbates hepatic lipid peroxidation and contributes to the development of steatohepatitis.[47] Ba exposure elicits hepatic oxidative stress, characterized by elevated malondialdehyde (MDA) levels and depleted glutathione, indicative of cellular damage.[48] Pb accumulates in the liver and activates inflammatory pathways through NF-κB signaling, promoting the release of pro-inflammatory cytokines.[35] W, by acting as a competitive inhibitor of molybdenum cofactors, disrupts essential hepatic enzymes such as sulfite oxidase and xanthine oxidase, impairing detoxification and purine metabolism.[45] While most prior studies have focused on the impact of individual metals, the combined and interactive effects of metal mixtures on human metabolic health remain poorly understood. Our findings suggest that Cd, Cs, Ba, Pb, and W may act synergistically or additively to contribute to NAFLD pathogenesis. The identification of these metals as robust predictors in our ML model not only provides empirical support for their involvement in NAFLD but also points to novel environmental risk factors and potential therapeutic targets for prevention and early intervention.\u003c/p\u003e\n\u003cp\u003eIn addition to the five key heavy metals, our study underscores the exceptional predictive value of body roundness index (BRI)—a novel anthropometric metric designed to more accurately capture regional adiposity compared to traditional measures such as BMI. Among the ten key predictors identified, BRI emerged as the single strongest contributor to NAFLD risk in the XGBoost model. By design, BRI quantifies abdominal adiposity and ectopic fat deposition, both of which are strongly linked to hepatic steatosis through insulin resistance, lipotoxicity, and chronic low-grade inflammation.[19] These mechanisms underpin the role of visceral fat—as morphologically manifested by a rounded body shape—in promoting hepatic lipid accumulation and metabolic dysfunction. The robust predictive power of BRI observed in our model supports the pathophysiological relevance of central obesity in NAFLD development. Further research exploring the molecular and metabolic pathways linking BRI to NAFLD could deepen our mechanistic understanding and potentially establish BRI as a clinically useful marker for early identification and stratification of at-risk individuals.\u003c/p\u003e\n\u003cp\u003eSeveral limitations of our study warrant consideration. First, the observational design of NHANES inherently limits our ability to establish causal relationships between urinary heavy metal exposure and NAFLD risk. Second, although machine learning models are well-suited for detecting complex, nonlinear interactions, the potential for residual confounding by unmeasured variables remains. ML-based inference should be complemented by mechanistic studies to confirm biological plausibility. Third, urinary heavy metals were measured at a single time point, which precludes evaluation of long-term or cumulative exposure—a key consideration in understanding environmental contributors to chronic liver disease. Fourth, because our analysis was conducted in a U.S.-based adult cohort, the generalizability of findings to populations with differing genetic, dietary, or sociocultural backgrounds remains uncertain. Future studies should pursue external validation in ethnically and geographically diverse cohorts. Fifth, while SHAP-based interpretability analyses enhance model transparency and highlight influential features, they do not substitute for toxicological or biological validation. Clinical studies are needed to determine threshold concentrations of heavy metals that confer meaningful NAFLD risk, and to elucidate the biological mechanisms underlying these associations.\u003c/p\u003e\n\u003cp\u003eIn this nationally representative cohort study, XGBoost demonstrated superior accuracy in predicting NAFLD compared to alternative machine learning models. Notably, a minimal feature set comprising five urinary heavy metals and five phenotypic indices achieved comparable predictive performance, with SHAP analyses providing robust interpretability and validation. To support clinical application, we developed a publicly accessible interactive web tool that generates individualized NAFLD risk profiles from routine clinical data. These findings highlight the promise of machine learning and urinary heavy metals as valuable tools in NAFLD risk stratification. Future efforts should focus on integrating interpretable ML models into clinical workflows to realize the full potential of AI-driven prevention strategies.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors would like to sincerely thank the NHANES participants and staff for contributing to the data collection and making the data available for public use.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding information\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the Public Service Development and Reform Pilot Project of Beijing Medical Research Institute (W. Niu), the Capital’s Funds for Health Improvement and Research (Grant Number: 2024-2-1133).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors have no relevant financial or non-financial interests to disclose.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor information\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAuthors and Affiliations\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eCapital Institute of Pediatrics, Chinese Academy of Medical Sciences \u0026amp; Peking Union Medical College, Beijing, China.\u003c/p\u003e\n\u003cp\u003eXiaoqian Wang\u003c/p\u003e\n\u003cp\u003eGraduate School, Beijing University of Chinese Medicine, Beijing, China.\u003c/p\u003e\n\u003cp\u003eMei Xue\u003c/p\u003e\n\u003cp\u003eChina-Japan Friendship Hospital, Beijing, China.\u003c/p\u003e\n\u003cp\u003eMei Xue\u003c/p\u003e\n\u003cp\u003ePrecision, University of Michigan, Ann Arbor, Michigan, USA.\u003c/p\u003e\n\u003cp\u003eHannah Chang\u003c/p\u003e\n\u003cp\u003eNortheast Forestry University, Harbin City, Heilongjiang Province, China.\u003c/p\u003e\n\u003cp\u003eBochun Wang\u003c/p\u003e\n\u003cp\u003eCenter for Evidence-Based Medicine, Capital Institute of Pediatrics, No.2 Yinghua East St., Chaoyang District, Beijing, 100029, China.\u003c/p\u003e\n\u003cp\u003eWenquan Niu\u003c/p\u003e\n\u003cp\u003eDepartment of Medicine, University of Pittsburgh School of Medicine, 200 Meyran Avenue, Suite 200, Pittsburgh, PA, USA\u003c/p\u003e\n\u003cp\u003eChung-Chou H Chang\u003c/p\u003e\n\u003cp\u003eDepartment of Biostatistics, University of Pittsburgh School of Public Health, 200 Meyran Avenue, Suite 200, Pittsburgh, PA, USA.\u003c/p\u003e\n\u003cp\u003eChung-Chou H Chang\u003c/p\u003e\n\u003cp\u003ePrecision Health Program, Department of Radiology, College of Human Medicine, Michigan State University, 766 Service Road, East Lansing, MI, USA.\u003c/p\u003e\n\u003cp\u003eXiaoqun Dong\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eContributions\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eXiaoqian Wang and Mei Xue designed the study and analyzed the data. Hannah Chang and Bochun Wang co-validated the analysis. The first draft was written by Xiaoqian Wang. Chung-Chou H Chang, Xiaoqun Dong, and Wenquan Niu were responsible for validating and revising the final content. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eCorresponding authors\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eCorrespondence to Chungchou H Chang, Xiaoqun Dong or Wenquan Niu.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data collected complied with the ethical guidelines established by the relevant institutions and/or national research councils, as well as the 1964 Declaration of Helsinki and its subsequent amendments or equivalent ethical standards. Informed consent was obtained from all individual participants included in the study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe National Health and Nutrition Examination Survey datasets are publicly available at the National Center for Health Statistics of the Center for Disease Control and Prevention\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e(https:// www. cdc. gov/ nchs/ nhanes/ index. htm).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003col\u003e\n\u003cli\u003eRiazi K, Azhari H, Charette JH, Underwood FE, King JA, Afshar EE, Swain MG, Congly SE, Kaplan GG, Shaheen AA: \u003cstrong\u003eThe prevalence and incidence of NAFLD worldwide: a systematic review and meta-analysis\u003c/strong\u003e. \u003cem\u003eLancet Gastroenterol Hepatol \u003c/em\u003e2022, \u003cstrong\u003e7\u003c/strong\u003e(9):851-861.\u003c/li\u003e\n\u003cli\u003eAmini-Salehi E, Letafatkar N, Norouzi N, Joukar F, Habibi A, Javid M, Sattari N, Khorasani M, Farahmand A, Tavakoli S\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eGlobal Prevalence of Nonalcoholic Fatty Liver Disease: An Updated Review Meta-Analysis comprising a Population of 78 million from 38 Countries\u003c/strong\u003e. \u003cem\u003eArch Med Res \u003c/em\u003e2024, \u003cstrong\u003e55\u003c/strong\u003e(6):103043.\u003c/li\u003e\n\u003cli\u003eByrne CD, Targher G: \u003cstrong\u003eNAFLD: a multisystem disease\u003c/strong\u003e. \u003cem\u003eJ Hepatol \u003c/em\u003e2015, \u003cstrong\u003e62\u003c/strong\u003e(1 Suppl):S47-64.\u003c/li\u003e\n\u003cli\u003eYounossi ZM: \u003cstrong\u003eNon-alcoholic fatty liver disease - A global public health perspective\u003c/strong\u003e. \u003cem\u003eJ Hepatol \u003c/em\u003e2019, \u003cstrong\u003e70\u003c/strong\u003e(3):531-544.\u003c/li\u003e\n\u003cli\u003eMcPherson S, Hardy T, Henderson E, Burt AD, Day CP, Anstee QM: \u003cstrong\u003eEvidence of NAFLD progression from steatosis to fibrosing-steatohepatitis using paired biopsies: implications for prognosis and clinical management\u003c/strong\u003e. \u003cem\u003eJ Hepatol \u003c/em\u003e2015, \u003cstrong\u003e62\u003c/strong\u003e(5):1148-1155.\u003c/li\u003e\n\u003cli\u003eLoomba R, Adams LA: \u003cstrong\u003eThe 20% Rule of NASH Progression: The Natural History of Advanced Fibrosis and Cirrhosis Caused by NASH\u003c/strong\u003e. \u003cem\u003eHepatology \u003c/em\u003e2019, \u003cstrong\u003e70\u003c/strong\u003e(6):1885-1888.\u003c/li\u003e\n\u003cli\u003eSheka AC, Adeyi O, Thompson J, Hameed B, Crawford PA, Ikramuddin S: \u003cstrong\u003eNonalcoholic Steatohepatitis: A Review\u003c/strong\u003e. \u003cem\u003eJAMA \u003c/em\u003e2020, \u003cstrong\u003e323\u003c/strong\u003e(12):1175-1183.\u003c/li\u003e\n\u003cli\u003eZhang H, Zhou XD, Shapiro MD, Lip GYH, Tilg H, Valenti L, Somers VK, Byrne CD, Targher G, Yang W\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eGlobal burden of metabolic diseases, 1990-2021\u003c/strong\u003e. \u003cem\u003eMetabolism \u003c/em\u003e2024, \u003cstrong\u003e160\u003c/strong\u003e:155999.\u003c/li\u003e\n\u003cli\u003eKim JJ, Kim YS, Kumar V: \u003cstrong\u003eHeavy metal toxicity: An update of chelating therapeutic strategies\u003c/strong\u003e. \u003cem\u003eJ Trace Elem Med Biol \u003c/em\u003e2019, \u003cstrong\u003e54\u003c/strong\u003e:226-231.\u003c/li\u003e\n\u003cli\u003eJ\u0026auml;rup L: \u003cstrong\u003eHazards of heavy metal contamination\u003c/strong\u003e. \u003cem\u003eBr Med Bull \u003c/em\u003e2003, \u003cstrong\u003e68\u003c/strong\u003e:167-182.\u003c/li\u003e\n\u003cli\u003eLi W, Huang G, Tang N, Lu P, Jiang L, Lv J, Qin Y, Lin Y, Xu F, Lei D: \u003cstrong\u003eEffects of heavy metal exposure on hypertension: A machine learning modeling approach\u003c/strong\u003e. \u003cem\u003eChemosphere \u003c/em\u003e2023, \u003cstrong\u003e337\u003c/strong\u003e:139435.\u003c/li\u003e\n\u003cli\u003eChang W, Li P: \u003cstrong\u003eCopper and Diabetes: Current Research and Prospect\u003c/strong\u003e. \u003cem\u003eMol Nutr Food Res \u003c/em\u003e2023, \u003cstrong\u003e67\u003c/strong\u003e(23):e2300468.\u003c/li\u003e\n\u003cli\u003eWang Y, Wang Y, Li R, Ni B, Chen R, Huang Y, Cheng R, Li P, Li H, Peng Y\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eLow-grade systemic inflammation links heavy metal exposures to mortality: A multi-metal inflammatory index approach\u003c/strong\u003e. \u003cem\u003eSci Total Environ \u003c/em\u003e2024, \u003cstrong\u003e947\u003c/strong\u003e:174537.\u003c/li\u003e\n\u003cli\u003eSadighara P, Abedini AH, Irshad N, Ghazi-Khansari M, Esrafili A, Yousefi M: \u003cstrong\u003eAssociation Between Non-alcoholic Fatty Liver Disease and Heavy Metal Exposure: a Systematic Review\u003c/strong\u003e. \u003cem\u003eBiol Trace Elem Res \u003c/em\u003e2023, \u003cstrong\u003e201\u003c/strong\u003e(12):5607-5615.\u003c/li\u003e\n\u003cli\u003eZhang D, Wu S, Lan Y, Chen S, Wang Y, Sun Y, Lu Y, Liao W, Wang L: \u003cstrong\u003eEssential metal mixtures exposure and NAFLD: A cohort-based case-control study in northern Chinese male adults\u003c/strong\u003e. \u003cem\u003eChemosphere \u003c/em\u003e2023, \u003cstrong\u003e339\u003c/strong\u003e:139598.\u003c/li\u003e\n\u003cli\u003eZhang D, Wu S, Lan Y, Chen S, Wang Y, Sun Y, Liao W, Wang L: \u003cstrong\u003eBlood manganese and nonalcoholic fatty liver disease: A cohort-based case-control study\u003c/strong\u003e. \u003cem\u003eChemosphere \u003c/em\u003e2022, \u003cstrong\u003e287\u003c/strong\u003e(Pt 4):132316.\u003c/li\u003e\n\u003cli\u003eVandenbroucke JP, von Elm E, Altman DG, Gotzsche PC, Mulrow CD, Pocock SJ, Poole C, Schlesselman JJ, Egger M, Initiative S: \u003cstrong\u003eStrengthening the Reporting of Observational Studies in Epidemiology (STROBE): explanation and elaboration\u003c/strong\u003e. \u003cem\u003ePLoS Med \u003c/em\u003e2007, \u003cstrong\u003e4\u003c/strong\u003e(10):e297.\u003c/li\u003e\n\u003cli\u003eHsu C, Caussy C, Imajo K, Chen J, Singh S, Kaulback K, Le MD, Hooker J, Tu X, Bettencourt R\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eMagnetic Resonance vs Transient Elastography Analysis of Patients With Nonalcoholic Fatty Liver Disease: A Systematic Review and Pooled Analysis of Individual Participants\u003c/strong\u003e. \u003cem\u003eClin Gastroenterol Hepatol \u003c/em\u003e2019, \u003cstrong\u003e17\u003c/strong\u003e(4):630-637 e638.\u003c/li\u003e\n\u003cli\u003eThomas DM, Bredlau C, Bosy-Westphal A, Mueller M, Shen W, Gallagher D, Maeda Y, McDougall A, Peterson CM, Ravussin E\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eRelationships between body roundness with body fat and visceral adipose tissue emerging from a new geometrical model\u003c/strong\u003e. \u003cem\u003eObesity (Silver Spring) \u003c/em\u003e2013, \u003cstrong\u003e21\u003c/strong\u003e(11):2264-2271.\u003c/li\u003e\n\u003cli\u003eZhang X, Ma N, Lin Q, Chen K, Zheng F, Wu J, Dong X, Niu W: \u003cstrong\u003eBody Roundness Index and All-Cause Mortality Among US Adults\u003c/strong\u003e. \u003cem\u003eJAMA Netw Open \u003c/em\u003e2024, \u003cstrong\u003e7\u003c/strong\u003e(6):e2415051.\u003c/li\u003e\n\u003cli\u003ede L\u0026eacute;dinghen V, Wong VW, Vergniol J, Wong GL, Foucher J, Chu SH, Le Bail B, Choi PC, Chermak F, Yiu KK\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eDiagnosis of liver fibrosis and cirrhosis using liver stiffness measurement: comparison between M and XL probe of FibroScan\u0026reg;\u003c/strong\u003e. \u003cem\u003eJ Hepatol \u003c/em\u003e2012, \u003cstrong\u003e56\u003c/strong\u003e(4):833-839.\u003c/li\u003e\n\u003cli\u003eElSayed NA, Aleppo G, Aroda VR, Bannuru RR, Brown FM, Bruemmer D, Collins BS, Hilliard ME, Isaacs D, Johnson EL\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003e2. Classification and Diagnosis of Diabetes: Standards of Care in Diabetes-2023\u003c/strong\u003e. \u003cem\u003eDiabetes Care \u003c/em\u003e2023, \u003cstrong\u003e46\u003c/strong\u003e(Suppl 1):S19-s40.\u003c/li\u003e\n\u003cli\u003eWhelton PK, Carey RM, Aronow WS, Casey DE, Jr., Collins KJ, Dennison Himmelfarb C, DePalma SM, Gidding S, Jamerson KA, Jones DW\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003e2017 ACC/AHA/AAPA/ABC/ACPM/AGS/APhA/ASH/ASPC/NMA/PCNA Guideline for the Prevention, Detection, Evaluation, and Management of High Blood Pressure in Adults: A Report of the American College of Cardiology/American Heart Association Task Force on Clinical Practice Guidelines\u003c/strong\u003e. \u003cem\u003eHypertension \u003c/em\u003e2018, \u003cstrong\u003e71\u003c/strong\u003e(6):e13-e115.\u003c/li\u003e\n\u003cli\u003eRadzi SFM, Karim MKA, Saripan MI, Rahman MAA, Isa INC, Ibahim MJ: \u003cstrong\u003eHyperparameter Tuning and Pipeline Optimization via Grid Search Method and Tree-Based AutoML in Breast Cancer Prediction\u003c/strong\u003e. \u003cem\u003eJ Pers Med \u003c/em\u003e2021, \u003cstrong\u003e11\u003c/strong\u003e(10).\u003c/li\u003e\n\u003cli\u003eWarkentin MT, Al-Sawaihey H, Lam S, Liu G, Diergaarde B, Yuan JM, Wilson DO, Atkar-Khattra S, Grant B, Brhane Y\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eRadiomics analysis to predict pulmonary nodule malignancy using machine learning approaches\u003c/strong\u003e. \u003cem\u003eThorax \u003c/em\u003e2024, \u003cstrong\u003e79\u003c/strong\u003e(4):307-315.\u003c/li\u003e\n\u003cli\u003eZhou J, Meng X, Deng L, Liu N: \u003cstrong\u003eNon-linear associations between metabolic syndrome and four typical heavy metals: Data from NHANES 2011-2018\u003c/strong\u003e. \u003cem\u003eChemosphere \u003c/em\u003e2022, \u003cstrong\u003e291\u003c/strong\u003e(Pt 2):132953.\u003c/li\u003e\n\u003cli\u003eCanaday FT, Georas SN, Croft DP: \u003cstrong\u003eExamining the impact of air pollution, climate change, and social determinants of health on asthma and environmental justice\u003c/strong\u003e. \u003cem\u003eCurr Opin Pulm Med \u003c/em\u003e2024, \u003cstrong\u003e30\u003c/strong\u003e(3):276-280.\u003c/li\u003e\n\u003cli\u003eLavezzi AM, Ramos-Molina B: \u003cstrong\u003eEnvironmental Exposure Science and Human Health\u003c/strong\u003e. \u003cem\u003eInt J Environ Res Public Health \u003c/em\u003e2023, \u003cstrong\u003e20\u003c/strong\u003e(10).\u003c/li\u003e\n\u003cli\u003eHassanein EHM, Althagafy HS, Baraka MA, Amin H: \u003cstrong\u003eHepatoprotective effects of diosmin: a narrative review\u003c/strong\u003e. \u003cem\u003eNaunyn Schmiedebergs Arch Pharmacol \u003c/em\u003e2025, \u003cstrong\u003e398\u003c/strong\u003e(1):279-295.\u003c/li\u003e\n\u003cli\u003ePan Z, Gong T, Liang P: \u003cstrong\u003eHeavy Metal Exposure and Cardiovascular Disease\u003c/strong\u003e. \u003cem\u003eCirc Res \u003c/em\u003e2024, \u003cstrong\u003e134\u003c/strong\u003e(9):1160-1178.\u003c/li\u003e\n\u003cli\u003eLamas GA, Bhatnagar A, Jones MR, Mann KK, Nasir K, Tellez-Plaza M, Ujueta F, Navas-Acien A: \u003cstrong\u003eContaminant Metals as Cardiovascular Risk Factors: A Scientific Statement From the American Heart Association\u003c/strong\u003e. \u003cem\u003eJ Am Heart Assoc \u003c/em\u003e2023, \u003cstrong\u003e12\u003c/strong\u003e(13):e029852.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eCould Your Mysterious Health Issues Be Heavy Metal Poisoning Symptoms? \u003c/strong\u003e[https://www.ccfmed.com/blog/heavymetals]\u003c/li\u003e\n\u003cli\u003eKoyama H, Kamogashira T, Yamasoba T: \u003cstrong\u003eHeavy Metal Exposure: Molecular Pathways, Clinical Implications, and Protective Strategies\u003c/strong\u003e. \u003cem\u003eAntioxidants (Basel) \u003c/em\u003e2024, \u003cstrong\u003e13\u003c/strong\u003e(1).\u003c/li\u003e\n\u003cli\u003eKarri V, Schuhmacher M, Kumar V: \u003cstrong\u003eHeavy metals (Pb, Cd, As and MeHg) as risk factors for cognitive dysfunction: A general review of metal mixture mechanism in brain\u003c/strong\u003e. \u003cem\u003eEnviron Toxicol Pharmacol \u003c/em\u003e2016, \u003cstrong\u003e48\u003c/strong\u003e:203-213.\u003c/li\u003e\n\u003cli\u003eThitame SN, Aher AA, Mopagar V: \u003cstrong\u003eThe Detoxifying Effects of Spirulina in Promoting Liver Health and Heavy Metal Elimination - A Review\u003c/strong\u003e. \u003cem\u003eJ Pharm Bioallied Sci \u003c/em\u003e2025, \u003cstrong\u003e17\u003c/strong\u003e(Suppl 1):S43-S45.\u003c/li\u003e\n\u003cli\u003eChen H, Tan H, Wan J, Zeng Y, Wang J, Wang H, Lu X: \u003cstrong\u003ePPAR-gamma signaling in nonalcoholic fatty liver disease: Pathogenesis and therapeutic targets\u003c/strong\u003e. \u003cem\u003ePharmacol Ther \u003c/em\u003e2023, \u003cstrong\u003e245\u003c/strong\u003e:108391.\u003c/li\u003e\n\u003cli\u003eAlruhaimi RS, Hassanein EHM, Alotaibi MF, Alzoghaibi MA, Abd El-Ghafar OAM, Mohammad MK, Alnasser SM, Mahmoud AM: \u003cstrong\u003eDiallyl Disulfide Mitigates Cadmium Hepatotoxicity by Attenuating Oxidative Stress and TLR-4/NF-kappaB Signaling and Upregulating PPARgamma\u003c/strong\u003e. \u003cem\u003eFront Biosci (Landmark Ed) \u003c/em\u003e2024, \u003cstrong\u003e29\u003c/strong\u003e(10):369.\u003c/li\u003e\n\u003cli\u003eMahmood S, Tanvir EM, Komarova T, Islam MN, Khatun M, Hossain MF, Ng JC, Whitfield KM, Hossain MS, Khalil MI\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eRelationships between growth indicators, liver and kidney function markers, and blood concentrations of essential and potentially toxic elements in environmentally exposed young children\u003c/strong\u003e. \u003cem\u003eInt J Hyg Environ Health \u003c/em\u003e2023, \u003cstrong\u003e253\u003c/strong\u003e:114237.\u003c/li\u003e\n\u003cli\u003eLittle BB, Reilly R, Walsh B, Vu GT: \u003cstrong\u003eCadmium Is Associated with Type 2 Diabetes in a Superfund Site Lead Smelter Community in Dallas, Texas\u003c/strong\u003e. \u003cem\u003eInt J Environ Res Public Health \u003c/em\u003e2020, \u003cstrong\u003e17\u003c/strong\u003e(12).\u003c/li\u003e\n\u003cli\u003ePan K, Xu J, Xu Y, Wang C, Yu J: \u003cstrong\u003eThe association between endocrine disrupting chemicals and nonalcoholic fatty liver disease: A systematic review and meta-analysis\u003c/strong\u003e. \u003cem\u003ePharmacol Res \u003c/em\u003e2024, \u003cstrong\u003e205\u003c/strong\u003e:107251.\u003c/li\u003e\n\u003cli\u003eZhang L, Wang R, Xue Q, Wang Y, Xu J, Wang C, Fang X, Gao S, Zhang H, Guo L: \u003cstrong\u003eBioinformatic Analysis for Exploring Target Genes and Molecular Mechanisms of Cadmium-Induced Nonalcoholic Fatty Liver Disease and Targeted Drug Prediction\u003c/strong\u003e. \u003cem\u003eJ Appl Toxicol \u003c/em\u003e2025, \u003cstrong\u003e45\u003c/strong\u003e(5):858-865.\u003c/li\u003e\n\u003cli\u003eDeo RC: \u003cstrong\u003eMachine Learning in Medicine\u003c/strong\u003e. \u003cem\u003eCirculation \u003c/em\u003e2015, \u003cstrong\u003e132\u003c/strong\u003e(20):1920-1930.\u003c/li\u003e\n\u003cli\u003eZhang J, Ma X, Zhang J, Sun D, Zhou X, Mi C, Wen H: \u003cstrong\u003eInsights into geospatial heterogeneity of landslide susceptibility based on the SHAP-XGBoost model\u003c/strong\u003e. \u003cem\u003eJ Environ Manage \u003c/em\u003e2023, \u003cstrong\u003e332\u003c/strong\u003e:117357.\u003c/li\u003e\n\u003cli\u003eDou B, Zhu Z, Merkurjev E, Ke L, Chen L, Jiang J, Zhu Y, Liu J, Zhang B, Wei GW: \u003cstrong\u003eMachine Learning Methods for Small Data Challenges in Molecular Science\u003c/strong\u003e. \u003cem\u003eChem Rev \u003c/em\u003e2023, \u003cstrong\u003e123\u003c/strong\u003e(13):8736-8780.\u003c/li\u003e\n\u003cli\u003eChristie IN, Theparambil SM, Braga A, Doronin M, Hosford PS, Brazhe A, Mascarenhas A, Nizari S, Hadjihambi A, Wells JA\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eAstrocytes produce nitric oxide via nitrite reduction in mitochondria to regulate cerebral blood flow during brain hypoxia\u003c/strong\u003e. \u003cem\u003eCell Rep \u003c/em\u003e2023, \u003cstrong\u003e42\u003c/strong\u003e(12):113514.\u003c/li\u003e\n\u003cli\u003eSun J, Chen Y, Wang T, Ali W, Ma Y, Yuan Y, Gu J, Bian J, Liu Z, Zou H: \u003cstrong\u003eCadmium promotes nonalcoholic fatty liver disease by inhibiting intercellular mitochondrial transfer\u003c/strong\u003e. \u003cem\u003eCell Mol Biol Lett \u003c/em\u003e2023, \u003cstrong\u003e28\u003c/strong\u003e(1):87.\u003c/li\u003e\n\u003cli\u003eRacine R, Grandcolas L, Blanchardon E, Gourmelon P, Veyssiere G, Souidi M: \u003cstrong\u003eHepatic cholesterol metabolism following a chronic ingestion of cesium-137 starting at fetal stage in rats\u003c/strong\u003e. \u003cem\u003eJ Radiat Res \u003c/em\u003e2010, \u003cstrong\u003e51\u003c/strong\u003e(1):37-45.\u003c/li\u003e\n\u003cli\u003eElwej A, Chaabane M, Ghorbel I, Chelly S, Boudawara T, Zeghal N: \u003cstrong\u003eEffects of barium graded doses on redox status, membrane bound ATPases and histomorphological aspect of the liver in adult rats\u003c/strong\u003e. \u003cem\u003eToxicol Mech Methods \u003c/em\u003e2017, \u003cstrong\u003e27\u003c/strong\u003e(9):677-686.\u003c/li\u003e\n \u003c/ol\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003e\u003cstrong\u003eTable 1.\u003c/strong\u003e Baseline characteristics of study participants\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCharacteristics\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eNormal controls\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e(N=701)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eNAFLD patients\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e(N=512)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u003cem\u003eP-values\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAge (years)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e18-39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e282 (40.23)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e107 (20.90)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e40-59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e195 (27.82)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e188 (36.72)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u0026ge;60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e224 (31.95)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e217 (42.38)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMen\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e334 (47.65)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e282 (55.08)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e0.011\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRace/ethnicity\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eMexican American\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e64 (9.13)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e108 (21.09)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eNon-Hispanic Black\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e213 (30.39)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e99 (19.34)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eNon-Hispanic White\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e228 (32.52)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e172 (33.59)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eOther\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e196 (27.96)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e133 (25.98)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMarital status\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eMarried\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e372 (56.53)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e327 (65.14)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eSeparated\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e145 (22.04)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e109 (21.71)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eNever married\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e141 (21.43)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e66 (13.15)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eEducation level\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e0.094\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eLess than high school\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e117 (17.78)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e111 (22.07)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eHigh school grad/GED\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e165 (25.08)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e134 (26.64)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eCollege or above\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e376 (57.14)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e258 (51.29)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIncome\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e0.345\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eLow\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e113 (18.77)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e81 (18.28)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eIntermediate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e243 (40.37)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e198 (44.70)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eHigh\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e246 (40.46)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e164 (37.02)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCigarette smoking\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e0.231\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eNo smoking\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e393 (56.06)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e298 (58.20)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eFormer smoking\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e169 (24.11)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e132 (25.78)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eCurrent smoking\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e139 (19.83)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e82 (16.02)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAlcohol drinking\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e0.979\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eNo drinking\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e610 (91.45)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e452 (91.50)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eEver drinking\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e57 (8.55)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e42 (8.50)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePhysical activity (times/week)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e325 (67.43)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e273 (79.82)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e1-2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e29 (6.02)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e13 (3.80)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e3-5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e98 (20.33)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e50 (14.62)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u0026gt;5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e30 (6.22)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e6 (1.75)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBRI\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e0-20%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e221 (31.53)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e22 (4.30)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e20-40%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e186 (26.53)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e57 (11.13)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e40-60%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e127 (18.12)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e115 (22.46)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e60-80%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e91 (12.98)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e152 (29.69)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e80-100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e76 (10.84)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e166 (32.42)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e76 (10.84)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e180 (35.16)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eNo\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e625 (89.16)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e332 (64.84)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHypertension\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e339 (51.36)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e335 (69.50)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eNo\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e321 (48.64)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e147 (30.50)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCVD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e0.061\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e41 (6.25)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e46 (9.18)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003eNo\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e615 (93.75)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e455 (90.82)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLSM (\u003c/strong\u003e\u003cstrong\u003ekPa)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u0026ge; 8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e37 (5.28)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e92 (17.97)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u0026lt; 8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e664 (94.72)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e420 (82.03)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eUrinary creatinine (g/L), mean (SD)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e0.061 \u0026plusmn; 0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e0.10 \u0026plusmn; 0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e0.023\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTG (mg/dL), mean (SD)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e85.16 \u0026plusmn; 50.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e130.89 \u0026plusmn; 71.27\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTC (mg/dL), mean (SD)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 25px;\"\u003e\n \u003cp\u003e104.29 \u0026plusmn; 33.94\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 23px;\"\u003e\n \u003cp\u003e112.12 \u0026plusmn; 36.64\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eAbbreviations:\u003c/strong\u003e BRI, body roundness index; CVD, cardiovascular disease; DM, diabetes mellitus; GED, general educational development; LSM, liver serum marker; NAFLD, nonalcoholic fatty liver disease; PIR, ratio of family income to poverty; TC, total cholesterol; TG, triglyceride.\u003c/p\u003e\n\u003cp\u003eData are expressed as number (percentage) unless otherwise indicated.\u003c/p\u003e\n\u003cp\u003e\u003csup\u003e*\u003c/sup\u003eFor Income, low refers to PIR within [0, 1], intermediate within (1, 3], and high \u0026gt;3.\u003c/p\u003e\n\u003cp\u003e\u003csup\u003e**\u003c/sup\u003e\u003cem\u003eP values\u0026nbsp;\u003c/em\u003ewere calculated using independent t test or Mann-Whitney U test for continuous variables and using \u0026chi;\u003csup\u003e2\u003c/sup\u003e test for categorical variables, if appropriate.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2.\u0026nbsp;\u003c/strong\u003eAverage 31 metrics derived from four machine learning models in validation groups of five imputed datasets\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"739\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePerformance metrics\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLightGBM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eNeural network\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSVM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eXGBoost\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.7282\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.7039\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.7126\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.7249\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eArea Under the ROC Curve\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.7937\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.7685\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.7648\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.7983\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eBalanced Accuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.7201\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.6942\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.6963\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.7183\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eBinary Brier Score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.1830\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.2033\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.1956\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.1804\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eClassification Error\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.2718\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.2961\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.2874\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.2751\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eDiagnostic Odds Ratio\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e6.8399\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e5.3448\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e5.8305\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e6.6370\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eF\u0026beta;-Score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.6742\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.6427\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.6348\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.6742\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eFalse Discovery Rate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.3201\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.3473\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.3183\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.3281\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eFalse Negatives\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e67.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e74.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e82.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e66.0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eFalse Negative Rate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.3314\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.3667\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.4059\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.3235\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eFalse Omission Rate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.2376\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.2605\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.2695\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.2360\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eFalse Positives\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e64.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e68.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e56.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e67.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eFalse Positive Rate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.2285\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.2448\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.2014\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.2399\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eLogarithmic Loss\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.5406\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.6041\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.5812\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.5346\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eMulticlass AUC Type 1 Pairwise\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.7937\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.7685\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.7648\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.7983\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eMulticlass AUC Type 1 Unweighted\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.7937\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.7685\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.7648\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.7983\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eMulticlass AUC Type N Pairwise\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.7937\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.7685\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.7648\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.7983\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eMulticlass AUC Type N Unweighted\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.7937\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.7685\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.7648\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.7983\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eMulticlass AUC Macro-Averaged\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.7937\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.7685\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.7648\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.7983\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eMulticlass Brier Score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.3660\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.4067\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.3912\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.3607\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eMatthews Correlation Coefficient\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.4412\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.3903\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.4023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.4363\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eNegative Predictive Value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.7624\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.7395\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.7305\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.7640\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003ePositive Predictive Value\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.6799\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.6527\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.6817\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.6719\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003ePrecision-Recall AUC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.6980\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.6903\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.6908\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.7011\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.6799\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.6527\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.6817\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.6719\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eSensitivity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.6686\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.6333\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.5941\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.6765\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eSpecificity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.7715\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.7552\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.7986\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.7601\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eTrue Negatives\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e216.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e212.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e224.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e213.6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eTrue Negative Rate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.7715\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.7552\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.7986\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.7601\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eTrue Positives\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e136.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e129.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e121.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e138.0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 245px;\"\u003e\n \u003cp\u003eTrue Positive Rate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 114px;\"\u003e\n \u003cp\u003e0.6686\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 136px;\"\u003e\n \u003cp\u003e0.6333\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\n \u003cp\u003e0.5941\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e0.6765\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eAbbreviations:\u0026nbsp;\u003c/strong\u003eAUC, area under the ROC; LightGBM, light gradient boosting machine; ROC, receiver operating characteristic; SVM, support vector machine; XGBoost, eXtreme gradient boosting.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Heavy metals, Non-alcoholic fatty liver disease, Machine learning, SHapley Additive exPlanati","lastPublishedDoi":"10.21203/rs.3.rs-7286245/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7286245/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eObjectives:\u003c/strong\u003eThis study aimed to develop and validate an explainable machine learning (ML) model to predict NAFLD based on urinary heavy metals and phenotypic indices.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods:\u003c/strong\u003eData were drawn from the NHANES 2017-2020. NAFLD was defined as a controlled attenuation parameter (CAP)≥274 dB/m. Urinary heavy metals were quantified by inductively coupled plasma mass spectrometry and normalized to urinary creatinine to account for dilution. Four ML algorithms (LightGBM, NNET, SVM, and XGBoost) were implemented. The dataset was split into training (60%) and validation (40%) sets.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults: \u003c/strong\u003eAmong 1,213 adults, 512 were classified with NAFLD and 701 as controls. XGBoost outperformed others, achieving superior performance (AUC=0.7983; Brier score=0.1804). Feature importance was assessed using SHapley Additive exPlanations (SHAP), identifying a minimal subset of 10 features that preserved model performance. The strongest predictors were: body roundness index, triglyceride, diabetes mellitus, sex, age, and urinary concentrations of cadmium, cesium, barium, lead, and tungsten. Both global and local SHAP interpretations validated these features' contributions. The optimized XGBoost model was deployed as a web application (https://wxqdepression.shinyapps.io/nafldapp/).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions:\u003c/strong\u003e XGBoost demonstrated superior performance in predicting NAFLD using a streamlined set of urinary heavy metals and phenotypic indicators. SHAP-based interpretability confirmed the relevance of this minimal feature set.\u003c/p\u003e","manuscriptTitle":"Explainable machine learning model incorporating urinary heavy metals to predict nonalcoholic fatty liver disease","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-18 07:41:01","doi":"10.21203/rs.3.rs-7286245/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"be4044b7-3ec9-44cb-b098-2536d20791f3","owner":[],"postedDate":"August 18th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-09-13T17:38:24+00:00","versionOfRecord":[],"versionCreatedAt":"2025-08-18 07:41:01","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7286245","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7286245","identity":"rs-7286245","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}