Visceral adiposity index as a predictor of metabolic dysfunction-associated steatotic liver disease: a cross-sectional study

Abstract

Background

The association between visceral adiposity index (VAI) and metabolic dysfunction-associated steatotic liver disease (MASLD) remains unestablished. Our study sought to investigate the potential relationship between VAI and MASLD risk.

Methods

This study employed data from the 2017-2018 National Health and Nutrition Examination Survey (NHANES). Weighted multivariable regression models, subgroup analyses, and machine learning algorithms were used to evaluate associations and predictive performance.

Results

Higher VAI tertiles correlated with increased MASLD risk (adjusted OR for T3 vs. T1: 7.08, 95% CI: 4.35-11.5; P-trend=0.003). Machine learning models demonstrated robust predictive accuracy, with random forest (AUC=0.869) and gradient boosting machine (AUC=0.868) outperforming non-invasive scores. However, lipid accumulation product (LAP, AUC=0.834) and fatty liver index (FLI, AUC=0.833) achieved superior diagnostic performance compared to VAI (AUC=0.736), while maintaining clinical interpretability through simplicity and routine parameter availability.

Conclusions

While VAI demonstrated significant positive associations with MASLD risk, non-invasive scores like LAP and FLI emerged as superior diagnostic tools, balancing accuracy with clinical practicality.

Introduction

Metabolic dysfunction-associated steatotic liver disease (MASLD), formerly termed nonalcoholic fatty liver disease (NAFLD), constitutes a significant global public health challenge [1]. Current epidemiological data indicate its prevalence exceeds 80 million adults in the United States alone, representing 25–30% of the population [2], with projections estimating nearly 100 million affected individuals worldwide by 2030 [3]. This escalating burden is principally attributable to modifiable risk factors including unhealthy dietary patterns, sedentary behaviors, and insufficient physical activity [4, 5]. MASLD imposes substantial socioeconomic burdens through its association with multisystem comorbidities such as hepatocellular carcinoma, obesity-related metabolic syndrome, hypertension-mediated cardiovascular disorders, chronic kidney disease progression, and extrahepatic malignancies [6, 7]. Early identification of high-risk populations through effective stratification enables timely therapeutic interventions, thereby mitigating progression to decompensated cirrhosis, hepatic failure, and associated mortality [8]. While histological evaluation via liver biopsy remains the diagnostic reference standard, procedural risks and sampling variability necessitate the development of reliable non-invasive diagnostic modalities [9]. Although vibration-controlled transient elastography currently represents the clinical benchmark for fibrosis staging [10], its restricted availability in resource-limited settings and cost-intensive nature underscore the critical demand for scalable, standardized diagnostic protocols. Such advancements are imperative for longitudinal monitoring of fibrotic evolution and implementation of risk-adapted management strategies.

The visceral adiposity index (VAI) gained prominence as a critical biomarker in cardiometabolic risk, MASLD, and various endocrine diseases [11,12,13]. Recent breakthroughs in adipokine biology have redefined the pathophysiological understanding of visceral adiposity, revealing adipose tissue as a central orchestrator of MASLD progression [14]. Adipose tissue has been identified as a critical mediator of MASLD progression through multifaceted mechanisms including insulin resistance modulation, chronic inflammation induction, and autophagic dysfunction [15]. This mechanistic insight establishes the VAI, a composite metric incorporating waist circumference, triglyceride levels, HDL cholesterol, and body mass index, as a potential surrogate marker for these molecular processes [16]. A longitudinal cohort study with a 4-year follow-up demonstrated that VAI independently predicts MASLD incidence through a distinct dose-response relationship, with elevated VAI scores correlating strongly with fibrosis severity [17]. A systematic review and meta-analysis demonstrated that the VAI serves as a significant predictor for NAFLD and non-alcoholic steatohepatitis (NASH), with elevated VAI levels strongly correlating with severe hepatic steatosis compared to simple steatosis in adult populations [18]. The index demonstrates significant associations with multiple metabolic derangements, positioning it as a valuable clinical instrument for elucidating the obesity-metabolic-liver axis [19]. MASLD represents a redefined clinical entity that emphasizes the interplay between hepatic steatosis and cardiometabolic risk factors, replacing the former NAFLD nomenclature to better reflect its systemic metabolic underpinnings. While the visceral adiposity index (VAI) has been studied in NAFLD cohorts, its role in MASLD remains underexplored, particularly within the context of updated diagnostic criteria that prioritize metabolic dysfunction as a central diagnostic pillar. MASLD’s broader metabolic focus may recalibrate the clinical interpretation of biomarkers like VAI. However, existing studies have predominantly evaluated VAI in isolation, lacking comparative analyses against other non-invasive scores (e.g., LAP, FLI). This gap limits actionable insights into VAI’s distinct value in MASLD risk stratification. Furthermore, the transition from NAFLD to MASLD necessitates the revalidation of biomarkers under the revised framework, as diagnostic thresholds and metabolic confounders may differentially influence predictive accuracy.

The primary objective of this study was to examine the association between the VAI and MASLD utilizing epidemiological data from the 2017-2018 cycle of the NHANES, with particular emphasis on benchmarking its diagnostic performance against established non-invasive scores to address a critical gap in comparative risk stratification strategies.

Methods

Study population

Data were sourced from the NHANES, which employs a stratified multistage probability sampling design. From the 2017-2018 NHANES cohort containing 9,254 initial participants, the study applied sequential exclusion criteria: (1) 3,306 cases with missing controlled attenuation parameter (CAP) data for hepatic steatosis detection; (2) 894 participants (age $\ge$ 18 years) ineligible for the MASLD diagnostic criteria. The final analytical sample comprised 5054 eligible participants, as detailed in Fig. 1.

Fig. 1

Flowchart of the study participants

Diagnostic criteria

The Visceral Adiposity Index (VAI) is calculated through gender-specific formulas. For male subjects, the formula is expressed as: VAI = [WC/(39.68 + 1.88 $\times$ BMI)] $\times$ (TG/1.03) $\times$ (1.31/HDL)For female subjects, the formula is: VAI = [WC/(36.58 + 1.89 $\times$ BMI)] $\times$ (TG/0.81) $\times$ (1.52/HDL) where WC represents waist circumference (cm), BMI denotes body mass index (kg/m²), TG indicates triglyceride levels (mmol/L), and HDL signifies high-density lipoprotein cholesterol (mmol/L) [16]. Several non-invasive scores were calculated, such as fibrosis-4 index(FIB-4) [20], NAFLD fibrosis score(NFS) [21], fatty liver index(FLI) [22], hepatic steatosis index(HSI) [23], triglycerides index(TyG index) [24], lipid accumulation product(LAP) [25]. The controlled attenuation parameter (CAP) integrated into the FibroScan system enables non-invasive quantification of hepatic steatosis by assessing ultrasound signal attenuation caused by lipid deposition in hepatocytes [26]. Transient elastography with a controlled attenuation parameter (CAP) has demonstrated high diagnostic accuracy for assessing hepatic steatosis, where a CAP threshold of $\ge$238 dB/m is widely validated as a diagnostic criterion in clinical practice [27, 28]. The diagnosis of MASLD requires the presence of hepatic steatosis along with at least one of the following cardiovascular metabolic criteria: 1) BMI $\ge$25 kg/m² ($\ge$23 kg/m² in Asian populations) or increased waist circumference (>94 cm in males or >80 cm in females); 2) fasting glucose $\ge$5.6 mmol/L, 2-hour postprandial glucose $\ge$7.8 mmol/L, HbA1c $\ge$5.7%, type 2 diabetes, or diabetes treatment; 3) blood pressure $\ge$130/85 mmHg or antihypertensive therapy; 4) plasma triglycerides $\ge$1.70 mmol/L or lipid-lowering treatment; or 5) plasma HDL-cholesterol ($\le$1.0 mmol/L in males or $\le$1.3 mmol/L in females) or lipid-lowering treatment [1].

Covariables

The following covariates were selected based on prior research evidence and clinical relevance, categorized into four domains: (1) demographic characteristics (age, sex, race, educational attainment, marital status, and family poverty-income ratio); (2) questionnaire-based factors (moderate activities, self-reported diabetes, hypertension diagnosis, and smoking status, defined as having smoked geq100 cigarettes in lifetime); (3) laboratory measurements (platelet, albumin, AST, GGT, ALT, glucose, serum high-density lipoprotein cholesterol and triglyceride levels); (4) clinical examination parameters (waist circumference, body mass index, and median CAP value).

Statistical analysis

All statistical analyses were conducted using R statistical software (version 4.4.2; https://www.R-project.org) to appropriately accommodate the complex survey design of the NHANES database. Nationally representative estimates were ensured through the rigorous application of sampling weights in all analytical procedures. Participants with missing CAP measurements were excluded, while the remaining missing data in covariates were imputed using the k-nearest neighbors (KNN) algorithm, resulting in a complete-case dataset for subsequent analysis. Continuous variables are presented as weighted means with 95% confidence intervals (CI), while categorical variables are expressed as weighted proportions (95% CI). Group comparisons across VAI tertiles employed survey-weighted chi-square tests for categorical variables and survey-weighted linear regression for continuous measures. Given the limitations of conventional regression methods in complex survey data, we implemented survey design-weighted multivariable logistic regression models (Models 1–3 with progressive covariate adjustment for demographic, socioeconomic, and metabolic confounders) based on regression-based covariate control methodology to evaluate the independent association between visceral adiposity index (VAI) and MASLD risk. Results are reported as adjusted odds ratio (OR) with 95% CIs. The comparative performance evaluation of eight machine learning models (RF, SVM, GLM, GBM, KNN, NNET, DT) implemented using a stratified 7:3 training-test split and seven clinical scores was conducted using receiver operating characteristic (ROC) curve analysis, with area under the curve (AUC) serving as the primary discrimination metric.

Results

Baseline characteristics

Weighted population estimates (Table 1) demonstrated statistically significant differences across the VAI tertile-stratified groups for all baseline demographic and clinical variables analyzed (all P<0.05), except for gender distribution and family income.

Table 1 Weighted demographic baseline characteristics of the participants

Correlation between MASLD and VAI

The weighted multivariate regression analysis demonstrated progressively adjusted associations across three models: Model 1 (crude VAI association), Model 2 (adjusted for education, gender, race, marital status, age, and family income), and Model 3 (further adjusted for hypertension and diabetes status). This staged adjustment approach was designed to disentangle visceral adiposity effects from socioeconomic influences and metabolic comorbidities (Table 2). Significant positive associations were observed in both minimally adjusted models (Model 1: OR=2.68, 95% CI 2.37–3.03, P<0.001; Model 2: OR=2.68, 95% CI 2.29–3.14, P<0.001), with the association remaining statistically significant (OR=2.54, 95% CI 2.06–3.12, P<0.001) following full adjustment for all covariates in the final model. This corresponded to a 154% elevated risk of MASLD development per unit increment in VAI. When analyzed categorically through tertile stratification (T1-T3), the risk estimates demonstrated a dose-response pattern: compared with the reference tertile (T1), participants in higher tertiles exhibited progressively increased risks (T2: OR=1.84, 9% CI 1.13–3.01, P=0.033; T3: OR=7.08, 95% CI 4.35–11.5, P=0.003). A significant positive trend was observed across ascending VAI tertiles (P-trend=0.003).

Table 2 Association between MASLD and VAI

Subgroup analysis

To comprehensively assess the reliability of the observed positive association and examine potential variations across demographic strata, stratified analyses were performed through subgroup categorization and interaction testing. The investigation encompassed demographic variables including gender, age, ethnicity, educational attainment, and lifestyle factors such as physical activity patterns, marital status, diabetes prevalence, hypertension status, and smoking history. Notably, participants with elevated VAI levels consistently exhibited significantly increased MASLD risk across all stratified subgroups (all OR > 1 and all P < 0.05, Supplementary Fig. 1). Ethnicity-specific stratification revealed the most pronounced association in non-Hispanic Black populations (OR=3.46, 95% CI:2.67–4.50), followed by Other Ethnicities (OR=2.71, 95% CI:2.08–3.52) and non-Hispanic White cohorts (OR=2.53, 95% CI:2.13–3.00). Importantly, interaction testing demonstrated no statistically significant effect modification across any demographic parameters (P>0.05 for all interaction terms), indicating the robustness of VAI-MASLD association irrespective of population characteristics.

Selection of variables

Baseline characteristics were included in LASSO regression analysis to identify significant predictors of MASLD development (Fig. 2). Through 10-fold cross-validation with the lambda.1 se criterion, eleven parameters demonstrated predictive value: VAI, marital status, hypertension status, diabetes status, gender, age, race, platelet, albumin, AST, and ALT.

Fig. 2

Lasso regression for variables of interest: A regression coefficient profile diagram. B cross-validation curve

Nomogram development for risk prediction

Upon identification of the variables of interest, we developed a predictive model and presented its visualization through a nomogram (Fig. 3). The final model for osteoarthritis risk assessment incorporated eleven key predictors: VAI, marital status, hypertension status, diabetes status, gender, age, race, platelet, albumin, AST and ALT. Each predictor was assigned a quantitative score through standardized computation, with cumulative scores subsequently calculated to derive total risk values. A vertical projection from the total score axis to the risk probability axis enabled direct interpretation of MASLD risk stratification, where elevated scores correlated with progressive risk escalation.

Fig. 3

Nomogram for predicting MASLD risk

Prediction model

Among eight machine learning algorithms (Fig. 4A), random forest (RF) achieved the highest predictive accuracy (AUC=0.869), closely followed by gradient boosting machine (GBM, 0.868) and generalized linear model (GLM, 0.852), while neural network(NNET) showed relatively lower performance (AUC=0.504). In a parallel evaluation of seven clinical scores (Fig. 4B), lipid accumulation product (LAP, AUC=0.834) and fatty liver index (FLI, 0.833) emerged as superior predictors, outperforming established indices including hepatic steatosis index (HSI, 0.799), visceral adiposity index (VAI, 0.736), TyG index (0.732), NFS (0.606) and FIB-4 (0.553). Notably, the top-performing biomarkers (LAP and FLI) approached the diagnostic capability of mid-tier machine learning models, while maintaining clinically implementable simplicity compared to complex algorithms.

Fig. 4

Prediction model evaluation: A Machine learning model evaluation using ROC curves. B Non-invasive liver fibrosis biomarker evaluation using ROC curves

Discussion

This cross-sectional study analyzed data from 5054 participants in the 2017-2018 NHANES survey, revealing a significant positive association between VAI and MASLD risk, with consistent findings across diverse demographic subgroups. A nomogram prediction model incorporating eleven clinical parameters: VAI, marital status, hypertension status, diabetes status, gender, age, race, platelet count, albumin, AST, and ALT. Machine learning algorithms, particularly random forest (AUC=0.869) and gradient boosting machine (AUC=0.868), outperformed traditional clinical scores, underscoring the potential of advanced modeling in MASLD risk stratification. Comparative analysis of non-invasive biomarkers further identified lipid accumulation product (LAP, AUC=0.834) and fatty liver index (FLI, AUC=0.833) as superior predictors to VAI (AUC=0.736). While VAI’s dose-response relationship with MASLD risk (OR=7.08) underscores its value as an etiological biomarker, its moderate diagnostic accuracy positions it as a supplementary tool.

The mechanism underlying the observed relationships was delineated. Visceral fat accumulation emerged as an independent predictor of MASLD progression, while subcutaneous fat expansion demonstrated a correlation with disease regression, underscoring the differential roles of adipose tissue depots in regulating hepatic steatosis pathways [29,30,31]. Kim et al. demonstrated a longitudinal association between increased visceral adipose tissue area and elevated NAFLD incidence risk during a 4-year follow-up period, with their findings further revealing that adipose distribution plays a more critical role in NAFLD pathogenesis than total fat content [29]. Previous researchers quantified abdominal adiposity through CT-derived measurements of adipose tissue areas, though this methodology presents limitations of prohibitive cost and technical complexity for implementation in resource-constrained primary care settings [32]. The integration of anthropometric measurements with lipid profile parameters in the VAI demonstrates enhanced cost-effectiveness and diagnostic utility, with comparative analyses revealing VAI’s superior predictive capacity for NAFLD detection relative to isolated waist circumference assessments [33]. Notably, the discriminatory capacity of our nomogram aligns closely with models incorporating Fibrosis-4 scores or liver stiffness measurements, supporting VAI’s role in resource-constrained settings [34]. These findings propose VAI as a viable screening tool for visceral fat estimation in normal-weight populations and economically disadvantaged regions.

Our findings align with prior studies demonstrating VAI’s association with MASLD across diverse populations, yet highlight critical differences in diagnostic performance among non-invasive biomarkers. For instance, a study reported FLI as the strongest predictor of MASLD (AUC=0.82) [22], consistent with our results where FLI (AUC=0.833). LAP has been recognized as a robust predictor of MASLD in prior studies (AUC=0.84) [35], a finding reinforced by our analysis, which yielded an AUC of 0.834. Notably, machine learning models (e.g., random forest, AUC=0.869) surpassed all clinical scores, underscoring the potential of algorithmic approaches to integrate complex variable interactions. Clinically, VAI’s integration into nomograms could guide risk stratification in primary care settings, particularly where advanced imaging or machine learning tools are unavailable. Despite superior AUCs, machine learning models face translational barriers due to algorithmic complexity and resource demands, whereas LAP or FLI provide immediate clinical utility through routine parameter availability.

However, several limitations warrant consideration in interpreting these findings. First, the cross-sectional design precludes the definitive establishment of causal relationships between VAI and MASLD. While our analysis demonstrated robust associations, the temporal sequence of VAI elevation preceding MASLD onset remains unverified. Second, critical confounding variables such as dietary patterns, known to influence both visceral adiposity dynamics and hepatic steatosis pathogenesis, were not comprehensively captured in the NHANES dataset. Similarly, genetic polymorphisms affecting lipid metabolism that modify MASLD susceptibility were not analyzed, potentially obscuring gene-environment interactions central to disease mechanisms. Additionally, while CAP was used to diagnose hepatic steatosis, this study did not evaluate the performance of VAI and other non-invasive scores (e.g., LAP, FLI) across stratified grades of hepatic steatosis severity or fibrosis stages using CAP and liver stiffness measurements (LSM). Such stratification could clarify whether these scores retain diagnostic accuracy in advanced MASLD subpopulations. Future longitudinal studies integrating multi-omics profiling and fibrosis severity grading via CAP and LSM are essential.

Conclusion

The VAI demonstrated a positive correlation with MASLD risk, and its integration with clinical parameters improved risk stratification. Non-invasive scores like LAP and FLI, with performance metrics rivaling machine learning algorithms, should be prioritized for clinical implementation due to their simplicity and reliance on universally available metabolic parameters.