Metabolic dysfunction-associated fatty liver disease (MAFLD) has emerged as the most prevalent chronic liver condition globally, driven by the escalating burden of metabolic dysregulation, chronic inflammation, and insulin resistance [1, 2]. With an estimated prevalence exceeding 30% among adults worldwide, MAFLD is now a leading contributor to cirrhosis, hepatocellular carcinoma, and all-cause mortality, posing a significant global public health challenge [3]. In China, the prevalence of MAFLD is rising at an alarming rate, placing increasing strain on healthcare systems and national resources [4]. Early identification and precise risk stratification of high-risk individuals are crucial for effective prevention and intervention. However, current predictive tools lack robust composite indices that simultaneously capture metabolic overload and chronic inflammatory status, thereby limiting the development of efficient screening and targeted management strategies.

Against this backdrop, remnant cholesterol (RC)—a triglyceride-rich component of atherogenic lipoproteins primarily comprising very-low-density lipoproteins (VLDL), intermediate-density lipoproteins (IDL), and chylomicron remnants—has garnered increasing attention in the field of metabolic disease research [5]. Robust evidence from large-scale cohort studies has established RC as an independent predictor of cardiovascular events, beyond the traditional lipid markers LDL-C and HDL-C [6], and has implicated it as a key contributor to the pathogenesis of atherosclerosis. Recently, attention has turned to the potential mechanistic role of RC in metabolic liver diseases. Studies based on the NHANES population have demonstrated a strong association between serum RC levels and both hepatic steatosis and fibrosis in individuals with nonalcoholic fatty liver disease (NAFLD), with RC outperforming conventional cholesterol measures in predicting liver stiffness, underscoring its potential clinical relevance in hepatic risk stratification [7]. C-reactive protein (CRP), one of the most widely used biomarkers of chronic low-grade inflammation, has long been recognized as a critical factor in the progression of NAFLD/MAFLD [8–10]. Chronic systemic inflammation, moreover, has been shown to exacerbate adipose tissue dysfunction, disrupt insulin signaling pathways, and amplify hepatic steatosis and hepatocyte apoptosis via the adipose-liver axis [11, 12].

Single biomarkers often fall short in capturing the complex interplay between metabolic dysfunction and chronic inflammation in multifactorial diseases. To address this limitation, the RCII—a composite metric integrating RC and CRP—has recently been proposed as a surrogate for the coupled metabolic–inflammatory axis. Emerging evidence supports the predictive utility of RCII across a spectrum of chronic conditions. For instance, studies based on NHANES and CHARLS cohorts identified RCII as an independent predictor of incident stroke, with a graded increase in 7-year stroke risk across RCII quartiles [13]. Similarly, another NHANES-based analysis demonstrated that RCII outperforms RC or CRP alone in predicting all-cause, cardiovascular, and cancer-related mortality, exhibiting robust dose–response associations [14].

Although the RCII has shown preliminary promise in predicting several chronic diseases, its prognostic value, mechanistic relevance, and generalizability across populations in the context of MAFLD and associated mortality remain poorly characterized. In particular, it is unclear whether RCII exhibits a nonlinear association with MAFLD risk, whether it serves as an independent predictor, and whether its effects are mediated through specific metabolic pathways such as fasting plasma glucose. To date, no comprehensive epidemiological evidence has systematically addressed these questions. Furthermore, conventional statistical models are inherently limited in capturing complex feature interactions and high-dimensional data structures, potentially obscuring key risk patterns in chronic disease prediction. In contrast, machine learning approaches have emerged as powerful tools in medical risk stratification, offering enhanced accuracy, robustness, and the ability to uncover latent risk determinants in large-scale, multi-variable datasets.

To address these gaps, we leveraged cross-sectional and longitudinal data from the 1999–2010 National Health and Nutrition Examination Survey (NHANES) to systematically evaluate the association between the RCII and both MAFLD risk and related mortality outcomes. We further investigated potential nonlinear exposure–response relationships and mediating metabolic pathways. In parallel, we employed Boruta-based feature selection and a suite of machine learning algorithms to develop predictive models for MAFLD, aiming to establish RCII as a novel integrative biomarker and to advance intelligent modeling strategies for the early detection of metabolic diseases.

MAFLD was diagnosed according to the latest international consensus, requiring evidence of hepatic steatosis (via imaging or biochemical indicators) in addition to at least one of the following three metabolic conditions: (1) overweight or obesity; (2) diagnosed type 2 diabetes mellitus (T2DM); or (3) evidence of metabolic dysregulation. Metabolic dysregulation was defined as meeting at least two of the following six criteria: (1) central obesity (waist circumference > 102 cm in men or > 88 cm in women); (2) elevated blood pressure (systolic ≥ 130 mmHg or diastolic ≥ 85 mmHg, or current use of antihypertensive medication); (3) hypertriglyceridemia (triglycerides > 1.70 mmol/L or on lipid-lowering therapy); (4) reduced high-density lipoprotein cholesterol (HDL-C < 1.0 mmol/L in men or < 1.3 mmol/L in women); (5) prediabetes, defined as fasting plasma glucose of 5.6–6.9 mmol/L, 2-hour postprandial glucose of 7.8–11.0 mmol/L, or HbA1c of 5.7%–6.4%; (6) elevated high-sensitivity C-reactive protein (hsCRP > 2 mg/L) [15].

In the absence of abdominal imaging and liver biopsy data, this study employed the fatty liver index (FLI) to determine the presence of hepatic steatosis. The FLI was calculated according to the following formula:

where TG is serum triglycerides (mg/dL), BMI is body mass index (kg/m²), GGT is γ-glutamyl transferase (U/L), and waist circumference is measured in centimeters. Participants were classified as having hepatic steatosis if the FLI was ≥ 60, as previously validated in epidemiological studies [16].

The RCII, serving as the primary exposure variable in this study, is a composite marker integrating metabolic and inflammatory burden. It was calculated as:

RCII = RC × CRP,

where residual cholesterol (RC) was estimated using the formula RC = TC − (HDL-C + LDL-C), with all values expressed in mg/dL [17].

Mortality outcomes were ascertained through linkage of the NHANES cohort with the National Death Index (NDI), with follow-up through December 31, 2019. The primary endpoints included all-cause mortality (death from any cause), cardiovascular mortality (defined using ICD-10 codes I00–I09, I11, I13, I20–I51, and I60–I69), and premature death (defined as death occurring before the age of 75). Cause-of-death classifications were based on the "ucode_leading" variable provided in the publicly available mortality files curated by the National Center for Health Statistics (NCHS), ensuring standardized and consistent endpoint determination [18].

In multivariable analyses, a comprehensive set of covariates was included to account for potential confounding factors, encompassing demographic characteristics, lifestyle behaviors, and comorbid conditions. Demographic variables comprised age, sex, educational attainment, marital status, and race/ethnicity. Behavioral factors included smoking status and alcohol consumption. Clinical comorbidities—namely diabetes, hypertension, and dyslipidemia—were identified based on self-reported physician diagnoses or corresponding biochemical criteria, as previously defined [13].

To characterize the shape of the association between the RCII and the risk of MAFLD and mortality, restricted cubic spline (RCS) functions were applied. RCII was modeled as a continuous variable within logistic regression frameworks for MAFLD prevalence and Cox proportional hazards models for mortality outcomes, allowing for the identification of potential non-linear dose–response relationships. Knot placement was determined based on the empirical distribution of RCII, and all models were adjusted for key covariates [18].

Subgroup analyses were conducted to assess the robustness of the associations between RCII and study outcomes across strata of key demographic and health-related variables. Participants were stratified by age (< 60 vs. ≥60 years), sex (male vs. female), body mass index (BMI < 30 vs. ≥30 kg/m²), marital status (unmarried/widowed vs. married/cohabiting), educational attainment (high school or above vs. below high school), smoking status (non-smoker vs. smoker), alcohol consumption (no vs. yes), and history of diabetes, coronary heart disease, stroke, angina, and cancer—resulting in 12 predefined subgroups. Within each stratum, three progressively adjusted multivariable models (Model 1–3) were constructed to evaluate the association between RCII and outcome variables, with hazard ratios (HRs) and corresponding 95% confidence intervals reported. Statistical interactions were tested by incorporating multiplicative interaction terms into the fully adjusted models, and P-values for interaction were calculated to assess effect modification [20].

Kaplan–Meier survival curves were generated to estimate all-cause mortality, cardiovascular mortality, and premature death across quartiles of RCII. Group differences were assessed using the log-rank test. To further quantify the association between RCII and mortality outcomes, Cox proportional hazards models were constructed with three levels of covariate adjustment. Model 1 was unadjusted. Model 2 adjusted for age, sex, race, and educational attainment. Model 3 was additionally adjusted for body mass index (BMI), smoking status, alcohol consumption, and diabetes status. Results were reported as hazard ratios (HRs) with corresponding 95% confidence intervals (CIs). All analyses were performed using R software, and statistical significance was defined as a two-sided P-value < 0.05 [21].

Leveraging data from the NHANES cohort, this study systematically evaluated the association between the RCII and multiple adverse health outcomes, including MAFLD, all-cause mortality, cardiovascular mortality, and premature death. RCII demonstrated a robust and independent positive association with each endpoint, even after multivariable adjustment. Mendelian randomization and mediation analyses further supported the mechanistic roles of HDL-C, CRP, and FPG in this relationship. Among several machine learning algorithms, the RF model achieved superior predictive performance. SHAP analysis corroborated RCII as a key predictor of MAFLD risk, underscoring its potential utility in precision risk stratification and early identification of high-risk individuals.

In recent years, multiple research groups have employed machine learning techniques to develop predictive models for MAFLD, yielding promising results. However, variations in model performance, feature selection, and target populations have been noted across studies. A study based on NHANES 2017–2020 data developed a model centered on the non-HDL to HDL cholesterol ratio (NHHR), where the XGBoost algorithm achieved an AUC of 0.828. While demonstrating reasonable predictive power, the model relied predominantly on conventional lipid parameters and featured limited variable diversity [27]. In a large-scale investigation involving over five million individuals in Northwestern China, LASSO regression was used for feature selection, and the CatBoost model attained an AUC of 0.862, highlighting the predictive relevance of age, BMI, triglycerides, and fasting glucose; however, the absence of inflammatory markers limited its comprehensiveness [28]. Another study integrated vibration-controlled transient elastography (VCTE) parameters to stratify MAFLD risk, with an RF model achieving an AUC of 0.80 in the validation set, mainly for delineating low-, intermediate-, and high-risk groups [29]. Additionally, leveraging long-term follow-up data from NHANES III, a recent study employed multiple machine learning models to predict all-cause mortality among MAFLD patients, with the Coxnet model reaching an AUC of 0.88 at the 25-year mark—underscoring the clinical potential for long-term prognostic assessment [30]. We constructed and compared five machine learning models based on routine clinical and laboratory parameters. The RF model demonstrated superior performance in the test set, with an AUC of 0.960 and an accuracy of 89.7%, comparable to or exceeding previously reported models. Notably, the inclusion of RCII—a novel composite biomarker reflecting both metabolic dysfunction and systemic inflammation—substantially enhanced model interpretability and adaptability. These findings support the utility of RCII-enhanced machine learning models as robust tools for early MAFLD detection and individualized risk stratification.

Previous studies have established that RC is closely associated with hepatic lipid dysregulation and serves as a predictor for MAFLD and its cardiovascular outcomes [31]. Likewise, CRP, including hs-CRP, has been widely used as a convenient marker of systemic inflammation and is strongly linked to increased MAFLD risk [8] However, reliance on single biomarkers often yields inconsistent predictive performance across different populations, limiting their clinical robustness. In contrast, RCII, by integrating metabolic and inflammatory dimensions, demonstrates superior consistency and stability in predicting both MAFLD and mortality outcomes. Compared with other metabolic or inflammatory indicators—such as the systemic immune-inflammation index (SII) [32], homeostatic model assessment of insulin resistance (HOMA-IR) [33], and low thyroid function status [34] RCII exhibits greater external validity and translational potential in diverse clinical settings.

Multiple large-scale prospective cohort studies have confirmed that MAFLD is independently associated with elevated risks of all-cause and cardiovascular mortality [35–37], particularly among individuals with coexisting diabetes or the "lean MAFLD" phenotype [38, 39]. Extending this evidence, our study further reveals a positive and independent association between elevated RCII levels and premature death. Notably, this relationship persists even after comprehensive adjustment for potential confounders. These findings suggest that RCII may serve not only as a general prognostic marker for mortality, but also as an early-warning indicator for premature death—offering critical value for optimizing the timing of interventions and informing public health resource allocation.

The predictive power of the RCII for MAFLD, all-cause mortality, and cardiovascular mortality likely stems from its integration of two central pathological axes: metabolic dysregulation and chronic inflammation. RCII combines RC, a marker of lipid accumulation, with C-reactive protein (CRP or hs-CRP), a canonical indicator of systemic inflammation—each representing distinct yet interrelated biological pathways implicated in the pathogenesis and progression of metabolic diseases [40, 41]. Their synergistic interaction may potentiate vascular injury and organ dysfunction across multiple systems. Elevated RC promotes lipid deposition within arterial walls, contributes to endothelial dysfunction, and exacerbates lipid derangements via impaired reverse cholesterol transport mediated by HDL-C [42, 43]. In parallel, CRP not only suppresses pancreatic β-cell function but also stimulates hepatic glucose production, thereby aggravating FBG levels [44]. In our mediation analysis, FBG emerged as a significant intermediary linking RCII to increased mortality risk. Notably, elevated FBG serves as both a surrogate for insulin resistance and a pathogenic factor in its own right. Through activation of the AGE–RAGE axis, hyperglycemia accelerates vascular stiffening and myocardial remodeling, impairs HDL function, and promotes LDL oxidation—collectively compounding the metabolic and inflammatory disturbances driven by RCII [45, 46]. This tripartite interaction among dyslipidemia, inflammation, and glucose imbalance constitutes a tightly coupled risk network. Thus, RCII may be conceptualized as an integrated biomarker of metabolic–inflammatory stress, with its close association with FBG highlighting the pivotal role of glucose dysregulation in mediating the adverse outcomes linked to RCII.

This study has several limitations. First, the construction of RCII relies on the availability of CRP and lipid measurements, which may restrict its applicability in populations lacking inflammatory biomarker data. Second, although both traditional regression models and multiple machine learning algorithms consistently supported the robustness of RCII in assessing MAFLD risk, the cross-sectional nature of the NHANES dataset precludes definitive causal inference. Potential reverse causality and residual confounding cannot be fully ruled out. While Mendelian randomization provided suggestive evidence for a causal role, its validity is inherently constrained by the choice of instrumental variables and the characteristics of the study population. Therefore, longitudinal cohort studies and interventional trials are needed to further validate the prognostic utility of RCII.

In summary, RCII—a novel composite biomarker integrating metabolic and inflammatory signals—demonstrates strong predictive capacity and cross-model stability for MAFLD risk assessment, underscoring its potential clinical relevance. Future investigations should prioritize validating the accuracy and clinical effectiveness of RCII through animal models, randomized trials, and prospective cohort studies. Moreover, integrating multi-omics approaches, such as transcriptomics and metabolomics, may help elucidate the underlying metabolic–inflammatory pathways captured by RCII and provide mechanistic insight to support its translation into clinical practice.

As a composite biomarker integrating lipid dysregulation and systemic inflammation, the RCII is independently and significantly associated with increased risk of MAFLD and adverse mortality outcomes, demonstrating strong predictive utility. Its underlying mechanism may be partially mediated by elevated FPG levels.