The albumin-to-creatinine ratio predicts and explores potential mediation of mortality in metabolic dysfunction-associated steatotic liver disease in U.S. adults: evidence from NHANES 1999–2018

Abstract

Background

Metabolic dysfunction-associated steatotic liver disease (MASLD) frequently coexists with chronic kidney disease, which may exacerbate adverse outcomes. The albumin-to-creatinine ratio (ACR), an established marker of renal damage, has been associated with mortality risk in this population, yet its prognostic value and potential role in related pathways remain unclear.

Methods

We used data from the 1999–2018 National Health and Nutrition Examination Survey (NHANES) to construct MASLD cohorts defined by three indices: the Fatty Liver Index (FLI), United States Fatty Liver Index (USFLI), and Hepatic Steatosis Index (HSI). Weighted Cox models, restricted cubic splines, subgroup and sensitivity analyses assessed associations between ACR and all-cause and cardiovascular mortality. Mediation analyses examined the extent to which ACR might partially account for the observed associations between diabetes, hypertension, and mortality. Machine learning models were applied to evaluate predictive performance and identify key mortality predictors.

Results

Higher ACR levels were significantly associated with increased risks of all-cause and cardiovascular mortality (P < 0.001). Mediation analysis suggested that ACR may partially account for 33%–49% of the association between diabetes and all-cause mortality, and 19%–25% of the association between hypertension and all-cause mortality, although sensitivity analyses indicated that unmeasured confounding could influence these estimates. In stratified analyses, mortality risks increased progressively with higher fibrosis-4 index stages. Machine learning analyses demonstrated robust predictive performance across models and cohorts, with age and ACR consistently ranked as top predictors. Results were consistent across MASLD definitions and robust in sensitivity analyses.

Conclusions

ACR independently predicts mortality in MASLD and may partly account for the associations of diabetes and hypertension with mortality. Machine learning analyses supported its role as a key predictor of all-cause mortality. Its association with outcomes remains robust across fibrosis stages, underscoring its utility as a non-invasive biomarker for clinical risk assessment.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12876-025-04440-7.

Introduction

Metabolic dysfunction-associated steatotic liver disease (MASLD) is a globally prevalent metabolic disorder closely associated with obesity, type 2 diabetes, hypertension, sarcopenia, and cardiovascular disease (CVD) [1, 2]. Previously known as non-alcoholic fatty liver disease (NAFLD), MASLD reflects a recent redefinition that highlights the metabolic origins of hepatic steatosis, as endorsed by an international Delphi consensus [3]. Emerging evidence suggests that MASLD represents not merely hepatic steatosis, but a manifestation of systemic metabolic dysfunction [4, 5]. Cardiovascular and renal complications are important contributors to mortality in individuals with MASLD [6, 7]. Microalbuminuria, a marker of glomerular dysfunction and systemic endothelial damage, reflects early renal impairment and is indicative of generalised inflammation and metabolic stress [8, 9]. The albumin-to-creatinine ratio (ACR) is a widely used clinical measure for detecting microalbuminuria and has emerged as a biomarker indicative of global cardiometabolic risk. Elevated ACR, even within the normal range, has been associated with increased risk of CVD, metabolic derangements, and mortality [10–12]. Moreover, ACR plays an established role in chronic kidney disease (CKD) diagnostics and is increasingly recognised as an integrative marker along the lipid-heart–kidney–metabolism axis [13]. Diabetes and hypertension, both common comorbidities of MASLD, promote multiorgan damage via mechanisms such as chronic inflammation, oxidative stress, and vascular dysfunction [14, 15]. In this context, ACR may mediate the association between these conditions and mortality, offering insight into potential mechanistic pathways and therapeutic targets.

However, evidence remains limited regarding the prognostic value of ACR across multiple validated MASLD definitions and its potential mechanistic role as a mediator linking cardiometabolic comorbidities to mortality. In addition, current evidence underscores that hepatic fibrosis—assessed by non-invasive methods—strongly predicts CKD development and mortality among NAFLD patients [16, 17]; yet the interrelationship between ACR, fibrosis status, and long-term outcomes has not been well characterized in the MASLD population. In this context, we used data from the National Health and Nutrition Examination Survey (NHANES), a nationally representative program that combines household interviews with standardized physical examinations and extensive laboratory testing, to investigate the relationship between ACR and mortality across three established MASLD definitions. We further investigated ACR’s mediating role in the pathways linking diabetes and hypertension to mortality. Understanding these relationships may improve systemic risk stratification and inform clinical application of ACR in managing high-risk metabolic liver disease.

Methods

Data source and study population

This study was based on the NHANES data from 1999 to 2018. NHANES is a nationally representative survey employing a multistage, stratified, probability sampling design. The study received ethical approval from the National Center for Health Statistics (NCHS) Ethics Review Board. Moreover, all participants furnished written informed consent. Exclusion criteria were: age < 18 years, pregnancy, missing demographic, lifestyle or anthropometric data; presence of infectious diseases (hepatitis B/C); excessive alcohol consumption [18]; known hepatotoxic medication use [19, 20]; missing key variables (e.g., ACR, glycosylated hemoglobin type A1c (HbA1c), lipid profile, comorbidities, inflammatory indices, Fibrosis-4 index (FIB-4) [21]); or incomplete mortality follow-up. MASLD was identified as the presence of hepatic steatosis, defined using non-invasive indices including the Fatty Liver Index (FLI) [22], United States Fatty Liver Index (USFLI) [23], and Hepatic Steatosis Index (HSI) [24], in combination with at least one feature of metabolic dysfunction, in line with the recent multisociety Delphi consensus statement [3]. Metabolic dysfunction features included: type 2 diabetes; hypertension; dyslipidaemia (triglycerides (TG) ≥ 150 mg/dL, high-density lipoprotein cholesterol ≤ 40 mg/dL in men or ≤ 50 mg/dL in women, or use of lipid-lowering therapy); obesity or elevated waist circumference (WC); or impaired fasting blood glucose (FBG)/HbA1c. Detailed definitions are provided in Supplementary Table S1.The final analytic sample comprised 4,312 participants (FLI ≥ 60), 3,115 (USFLI ≥ 30), and 5,753 (HSI ≥ 36). The study flowchart is shown in Fig. 1. To clarify the overlap and distinctions among different MASLD definitions, we illustrated the distribution of participants identified by the FLI, USFLI, and HSI indices using a Venn diagram (Figure S1). Additional NHANES data are available at: https://www.cdc.gov/nchs/nhanes/.

Fig. 1.

Flowchart of participant selection in the MASLD study

Exposure and outcome definitions

Urinary albumin and creatinine data were obtained from the NHANES database according to the corresponding variable codes for different survey cycles. Specifically, urinary albumin was recorded under the code “URXUMASI” during the 1999–2004 cycles, and under “URXUMS” from 2005 to 2018. Urinary creatinine was consistently coded as “URXUCR” throughout all cycles, although the measurement method changed from the Jaffé method (1999–2006) to an enzymatic method (2007 onwards). According to NHANES technical documentation, this change introduced only minimal bias (average − 0.77%), and piecewise calibration equations were provided accordingly. In our primary analyses, ACR was calculated directly from the available urinary albumin and creatinine values [25]. To test robustness, we further conducted sensitivity analyses in which urinary creatinine values from pre-2007 cycles were calibrated using the recommended equations, the corresponding recalculated ACR values were applied, and survey cycle was additionally adjusted as a covariate. ACR was calculated as the ratio of urinary albumin (mg/L) to creatinine (g/L), and baseline values were used for analysis. Proteinuria was defined as ACR ≥ 30 mg/g, per Kidney Disease: Improving Global Outcomes (KDIGO) guidelines [26]. Primary outcomes were all-cause and cardiovascular mortality, ascertained via linkage to the National Death Index through 31 December 2019 [27]. Cardiovascular deaths were classified according to the International Classification of Diseases, 10th Revision (ICD-10), including codes I00–I09, I11, I13, I20–I51 for heart disease, and I60–I69 for cerebrovascular disease.

Covariates and confounders

We adjusted for demographic, socioeconomic, lifestyle, physical parameters, laboratory indicators and chronic disease status. Demographics included age, sex, race/ethnicity, marital status. Socioeconomic variables included education and income. Lifestyle factors included smoking and drinking. Comorbidities comprised diabetes, hypertension, CVD, and metabolic syndrome (MetS). Diabetes was defined as FBG ≥ 126 mg/dL, HbA1c ≥ 6.5%, a self-reported physician diagnosis, or current antidiabetic therapy. Hypertension was defined as systolic blood pressure (SBP) ≥ 130 mmHg, diastolic blood pressure (DBP) ≥ 80 mmHg (average of three measurements), or self-reported diagnosis. CVD included self-reported coronary heart disease, angina, myocardial infarction, or stroke. MetS was defined according to the National Cholesterol Education Program Adult Treatment Panel III (NCEP ATP III) criteria [28]. Anthropometric and laboratory variables included body mass index (BMI), WC, FBG, HbA1c, total cholesterol (TC), TG, low-density lipoprotein cholesterol (LDL-C), aspartate aminotransferase (AST), alanine aminotransferase (ALT), γ-glutamyl transpeptidase (GGT), serum urea (UREA), creatinine (CREA), uric acid (UA), estimated glomerular filtration rate (eGFR) [29]. Liver fibrosis was assessed using the FIB-4, a validated non-invasive score calculated from age, AST, ALT, and platelet count [21]; higher FIB-4 values indicate greater degrees of fibrosis. Inflammatory markers included neutrophil and lymphocyte counts, which were used to calculate the neutrophil-to-lymphocyte ratio (NLR), the systemic inflammation response index (SIRI) [30], and the systemic immune-inflammation index (SII) [31]. To address missing data, we adopted a complete-case approach: participants with missing values in any covariates required for a given model were excluded.

Statistical analysis

Analyses were performed in R version 4.4.3, accounting for the complex NHANES survey design, including sample weights (WTMEC2YR), stratification (SDMVSTRA), and clustering (SDMVPSU), in accordance with NHANES analytic guidelines for combining multiple cycles [32]. Continuous variables were reported as weighted means ± standard deviations (SD) or medians [interquartile ranges (IQR)]; categorical variables were reported unweighted counts and weighted percentages. The normality of continuous variables was assessed using the Shapiro–Wilk test. Group comparisons accounted for the complex NHANES survey design (weights, stratification, and clustering). Continuous variables were compared using survey-weighted t-tests or design-based Mann–Whitney U tests (two groups) and Kruskal–Wallis tests (multiple groups), as appropriate. Categorical variables were compared using Rao–Scott χ² tests. Survival differences by ACR category (< 30 vs. ≥30 mg/g) were assessed using Kaplan–Meier curves and log-rank tests in all three MASLD cohorts. Restricted cubic spline (RCS) models with three knots placed at the 10th, 50th, and 90th percentiles of the natural logarithm of ACR (log[ACR]) were fitted to flexibly assess potential nonlinear associations. Four survival models—random survival forest (RSF), CoxBoost, gradient boosting machine (GBM), and least absolute shrinkage and selection operator (Lasso) penalized Cox regression (Lasso-Cox)—were constructed using a 70/30 training-validation split. Boruta feature selection was applied to reduce dimensionality before modelling. RSF models were fitted using 1,000 trees with the log-rank splitting rule. The number of candidate variables randomly sampled at each split (mtry) was set to the default √p (where p is the number of predictors), and the minimum terminal node size was set to 15, the recommended default for survival outcomes in randomForestSRC. Bootstrap sampling without replacement was applied, and variable importance was computed to rank predictors. The CoxBoost model was implemented with 200 boosting steps (stepno = 200), while the penalty parameter was data-adaptively determined using the optimCoxBoostPenalty function to achieve optimal model regularization and prevent overfitting. The GBM model was fitted with 500 boosting iterations, a maximum depth of 3, and a shrinkage rate of 0.01, parameters chosen to balance computational efficiency and overfitting risk in moderately sized datasets. The Lasso-Cox was trained with the penalty parameter selected via 10-fold cross-validation. Model performance was evaluated using Harrell’s concordance index (C-index), time-dependent area under the curve (AUC), and Brier scores, while variable importance rankings were used to identify key predictors. In addition, calibration curves, decision curve analyses (DCA), integrated discrimination improvement (IDI), and net reclassification improvement (NRI) were conducted to assess model calibration, clinical utility, and potential improvements in predictive performance. Variable importance was assessed using model-specific metrics: permutation-based variable importance (VIMP) for RSF, relative influence for GBM, and the absolute values of regression coefficients for CoxBoost and Lasso-Cox. Variable importance rankings were then used to identify key predictors of all-cause mortality. Mediation analyses were performed using the mediation R package to assess ACR’s potential indirect role in the associations of diabetes and hypertension with all-cause and cardiovascular mortality. Primary models included a minimal set of covariates to reduce confounding while avoiding adjustment for variables potentially on the causal pathway (e.g., eGFR, FIB-4, lipids). Nonparametric bootstrap with 1,000 replications estimated 95% confidence intervals (CIs) for total, direct, and indirect effects, as well as proportion mediated. Sensitivity analyses using the medsens function added clinical covariates to examine the robustness of the mediation estimates to potential residual confounding. Subgroup analyses were conducted by age, sex, smoking status, drinking status, eGFR, ALT level, metabolic syndrome, diabetes, and hypertension, with interaction terms tested. Sensitivity analyses included recalculating ACR using calibration equations, excluding individuals who died within two years of follow-up, removing those with baseline malignancies, additionally adjusting for the use of renin–angiotensin–aldosterone system (RAAS) blockers and statins, re-evaluating associations using ACR quartiles to assess the robustness of the dose–response relationship, and performing competing risk analyses for cardiovascular mortality [33]. Primary Cox proportional hazards models incorporated survey weights, strata, and clustering, whereas machine-learning survival models were unweighted due to the absence of established approaches for complex survey adjustment and were interpreted as exploratory analyses. All models were sequentially adjusted for demographic, socioeconomic, lifestyle, comorbidity, and laboratory variables. Potential confounders were identified based on prior knowledge and cohort characteristics using a directed acyclic graph (DAG). Multicollinearity among covariates was evaluated with Tolerance and Variance Inflation Factor (VIF) using the olsrr R package. Two-sided P-values < 0.05 were considered statistically significant.

Results

Baseline characteristics

Table 1 summarises the weighted baseline characteristics of participants by ACR categories in the FLI-defined MASLD cohort. Compared to those with normal ACR levels (< 30 mg/g), individuals with elevated ACR were generally older, had lower eGFR, and showed higher levels of UA, TG, NLR, and SII, indicating greater inflammatory and metabolic burden. Additionally, the high ACR group had significantly higher prevalence of diabetes, hypertension, CVD, and MetS. In terms of population-level prevalence, 13.7% of participants in the FLI-defined cohort had ACR ≥ 30 mg/g, compared with 16.1% in the USFLI-defined cohort and 11.4% in the HSI-defined cohort. No significant differences were observed in sex, CREA, AST, or TC across ACR groups. These trends were consistently replicated in the USFLI- and HSI-defined MASLD cohorts. To evaluate the consistency among different MASLD definitions, agreement between FLI-, USFLI-, and HSI-based classifications was assessed using Cohen’s Kappa statistic (Supplementary Table S3). Agreement varied across definitions, with the strongest concordance observed between FLI and USFLI (Kappa = 0.30). For reference, detailed baseline characteristics across all three cohorts are provided in Supplementary Table S2. To further evaluate potential selection bias, we also summarised the patterns of missing data for each MASLD definition (Supplementary Tables S4–S6) and compared baseline characteristics between included and excluded participants (Supplementary Table S7).

Table 1.

Baseline characteristics by urinary ACR level in the FLI cohort

Characteristics	ACR < 30 mg/g N = 17,355,4701	ACR ≥ 30 mg/g N = 1,972,4591	Test of significance2
Sex, n (%)			χ²=2.06, p = 0.153
Male	2,111 (59%)	321 (54%)
Female	1,612 (41%)	268 (46%)
Age (years)	45.1 ± 14.5	52.9 ± 16.1	U = 7.9, p < 0.001
Smoking status, n (%)			χ²=3.08, p = 0.048
Never	1,964 (53%)	280 (46%)
Former	1,028 (27%)	190 (32%)
Current	731 (20%)	119 (22%)
Drinking status, n (%)			χ²=7.62, p < 0.001
Never	2,208 (62%)	286 (54%)
Former	744 (18%)	164 (25%)
Current	771 (20%)	139 (21%)
eGFR (mL/min/1.73 m²)	106.4 ± 11.9	101.1 ± 14.8	U=−5.69, p < 0.001
Log(ACR)	1.7 [1.4–2.2]	4.3 [3.7–5.1]	U = 77.63, p < 0.001
ACR (mg/g)	5.6 [3.9–8.8]	71.1 [41.6–171.0]	U = 77.63, p < 0.001
ALT (U/L)	26.0 [20.0–35.0]	24.0 [18.0–34.0]	U=−3.01, p = 0.003
AST (U/L)	23.0 [20.0–28.0]	23.0 [19.0–28.0]	U=−0.87, p = 0.388
UREA (mg/dL)	13.0 [10.0–16.0]	14.0 [11.0–18.0]	U = 2.71, p = 0.008
CREA (mg/dL)	0.9 [0.7–1.0.7.0]	0.8 [0.7–1.0.7.0]	U=−0.41, p = 0.682
UA (mg/dL)	5.9 [5.1–6.8]	6.1 [5.2–7.3]	U = 2.8, p = 0.006
HbA1c (%)	5.4 [5.2–5.6]	5.7 [5.4–7.3]	U = 9.07, p < 0.001
FBG (mg/dL)	99.0 [94.0–107.8.0.8]	110.9 [97.8–154.0]	U = 8.82, p < 0.001
TC (mg/dL)	205.6 ± 41.0	203.6 ± 43.5	U=−0.77, p = 0.442
TG (mg/dL)	145.0 [103.0–205.0.0.0]	150.0 [113.0–226.0.0.0]	U = 2.33, p = 0.021
NLR	2.1 ± 0.9	2.5 ± 1.4	U = 5.7, p < 0.001
SII	552.8 ± 349.2	636.2 ± 373.3	U = 4.76, p < 0.001
SIRI	1.0 [0.7–1.4]	1.2 [0.8–1.8]	U = 6.21, p < 0.001
FIB-4			χ²=25.35, p < 0.001
Low risk (F0-F2)	77 (2%)	44 (6%)
Indeterminate	716 (16%)	180 (26%)
High risk (F3-F4)	2,930 (83%)	365 (68%)
MetS, n (%)	2,136 (57%)	448 (72%)	χ²=23.9, p < 0.001
Hypertension, n (%)	1,550 (39%)	411 (62%)	χ²=67.09, p < 0.001
Diabetes, n (%)	749 (15%)	327 (48%)	χ²=176.98, p < 0.001
CVD, n (%)	277 (6%)	128 (18%)	χ²=95.21, p < 0.001
All-cause mortality, n (%)	439 (9%)	200 (30%)	U = 7.67, p < 0.001
CVD mortality, n (%)	121 (2%)	77 (11%)	U = 6.14, p < 0.001
Follow-up time (months)	126.0 [67.0- 186.0]	93.0 [48.0–160.0.0.0]	U=−5.05, p < 0.001
Weighted person-time (10⁶ person-years)	182.5	17.5	—

Abbreviations: ACR albumin-to-creatinine ratio, FLI Fatty Liver Index, eGFR estimated glomerular filtration rate, ALT alanine aminotransferase, AST aspartate aminotransferase, UREA serum urea, CREA creatinine, UA uric acid, HbA1c glycosylated hemoglobin type A1c, FBG fasting blood glucose, TC total cholesterol, TG triglyceride, NLR neutrophil-to-lymphocyte ratio, SII systemic immune-inflammation index, SIRI systemic Inflammation response index, FIB-4 fibrosis-4 index, MetS metabolic syndrome, CVD cardiovascular disease

¹ Weighted N = population estimate; n = unweighted sample size. Categorical variables are presented as unweighted n and weighted %. Continuous variables are presented as weighted mean ± SD or weighted median [Q1–Q3]

² Test of significance values are presented as the test statistic with corresponding p-values. χ² = Rao & Scott adjusted chi-square test; T = survey-weighted t-test; U = design-based Mann–Whitney U test

Association between ACR and mortality

In the three MASLD-defined cohorts, a total of 639 all-cause deaths and 198 cardiovascular deaths were observed in the FLI cohort, with a weighted median follow-up of 122 months. In the USFLI cohort, 548 all-cause and 211 cardiovascular deaths were observed, with a weighted median follow-up of 119 months. The HSI cohort recorded 649 all-cause deaths and 211 cardiovascular deaths, with a weighted median follow-up of 122 months. Kaplan–Meier survival analyses revealed significant differences in both all-cause and cardiovascular mortality across ACR categories, with consistent patterns observed across all three MASLD definitions (Fig. 2).

Fig. 2.

Kaplan–Meier curves of all-cause and cardiovascular mortality by ACR categories in MASLD cohorts. (Panels A–C)show all-cause mortality, and (Panels D–F)show cardiovascular mortality for participants with ACR < 30 mg/g vs. ≥30 mg/g in the three MASLD cohorts defined by FLI (A, D), USFLI (B, E), and HSI (C, F)

Table 2 presents the associations between ACR levels and the risk of all-cause and cardiovascular mortality in MASLD populations. After progressive adjustment for potential confounders, elevated ACR remained significantly associated with increased risk of both all-cause and cardiovascular mortality. These consistent results across all three MASLD definitions suggest that elevated ACR is an independent risk factor for mortality in this population.

Table 2.

Survey weight–adjusted multivariable Cox models for mortality by ACR in MASLD cohorts

Cohort	Exposure	Model 1 h (95% CI)	P value	Model 2 h (95% CI)	P value	Model 3 h (95% CI)	P value
A. All-cause Mortality
FLI	log(ACR)	1.57 (1.50–1.65)	< 0.001	1.41 (1.34–1.48)	< 0.001	1.28 (1.19–1.37)	< 0.001
	ACR ≥ 30 mg/g	4.23 (3.28–5.46)	< 0.001	2.78 (2.21–3.49)	< 0.001	2.12 (1.67–2.69)	< 0.001
USFLI	log(ACR)	1.56 (1.47–1.65)	< 0.001	1.44 (1.35–1.54)	< 0.001	1.33 (1.23–1.45)	< 0.001
	ACR ≥ 30 mg/g	3.93 (3.00–5.16)	< 0.001	3.02 (2.39–3.83)	< 0.001	2.39 (1.86–3.07)	< 0.001
HSI	log(ACR)	1.59 (1.52–1.66)	< 0.001	1.39 (1.31–1.47)	< 0.001	1.26 (1.18–1.35)	< 0.001
	ACR ≥ 30 mg/g	4.25 (3.35–5.39)	< 0.001	2.71 (2.12–3.47)	< 0.001	2.07 (1.58–2.70)	< 0.001
B. Cardiovascular Mortality
FLI	log(ACR)	1.68 (1.54–1.84)	< 0.001	1.53 (1.40–1.67)	< 0.001	1.38 (1.24–1.54)	< 0.001
	ACR ≥ 30 mg/g	5.86 (4.08–8.41)	< 0.001	3.75 (2.74–5.13)	< 0.001	2.52 (1.81–3.50)	< 0.001
USFLI	log(ACR)	1.60 (1.43–1.78)	< 0.001	1.48 (1.30–1.67)	< 0.001	1.37 (1.18–1.58)	< 0.001
	ACR ≥ 30 mg/g	4.95 (3.28–7.46)	< 0.001	3.56 (2.46–5.14)	< 0.001	2.67 (1.79–3.98)	< 0.001
HSI	log(ACR)	1.65 (1.51–1.81)	< 0.001	1.45 (1.31–1.60)	< 0.001	1.32 (1.18–1.49)	< 0.001
	ACR ≥ 30 mg/g	5.26 (3.61–7.65)	< 0.001	2.99 (2.11–4.23)	< 0.001	2.23 (1.56–3.18)	< 0.001

Abbreviations: ACR albumin-to-creatinine ratio, HR hazard ratio, CI confidence interval, FLI Fatty Liver Index, USFLI United States Fatty Liver Index, HSI Hepatic Steatosis Index, MASLD metabolic dysfunction-associated steatotic liver disease

Model 1 was unadjusted

Model 2 was adjusted for sex, age, race, poverty income ratio (PIR), education level, marital status, smoking status and drinking status

Model 3 was further adjusted for estimated glomerular filtration rate (eGFR), alanine aminotransferase (ALT), aspartate aminotransferase (AST), fibrosis-4 index (FIB-4), serum urea (UREA), creatinine (CREA), uric acid (UA), total cholesterol (TC), triglyceride (TG), neutrophil-to-lymphocyte ratio (NLR), systemic immune-inflammation index (SII), systemic Inflammation response index (SIRI), metabolic syndrome (MetS), hypertension, diabetes, and cardiovascular disease (CVD) in addition to covariates in Model 2

Furthermore, we examined the potential nonlinear association between log(ACR) and mortality risk (Fig. 3). The results indicated that, except for a nonlinear relationship between log(ACR) and all-cause mortality in the FLI cohort, the associations between log(ACR) and both all-cause and cardiovascular mortality were largely linear in the USFLI- and HSI-defined cohorts. These findings further reinforce the robust association between elevated ACR levels and adverse outcomes. Consistent results were observed in the weighted RCS analyses (Figure S2), showing significant overall and nonlinear effects. Although the curves appeared slightly U-shaped at the extremes due to sparse data and wide confidence intervals, the central range clearly showed a rising trend in mortality risk with higher ACR levels, consistent with the unweighted analyses.

Fig. 3.

Dose–response associations between log(ACR) and mortality in MASLD cohorts. RCS models show associations with all-cause mortality (Panels A–C) and cardiovascular mortality (Panels D–F) in the FLI-, USFLI-, and HSI-defined cohorts. Models were adjusted for covariates in Model 3. Solid lines represent hazard ratios (HRs), and shaded areas indicate 95% CIs

Subgroup and sensitivity analyses of ACR

To further explore the potential heterogeneity in the association between elevated ACR (≥ 30 mg/g) and mortality risk, subgroup analyses were conducted across the three MASLD-defined cohorts. For all-cause mortality, elevated ACR was consistently associated with significantly increased mortality risk in most subgroups, with detailed results provided in Supplementary Table S8. Although no statistically significant interactions were observed (all P for interaction > 0.05), relatively higher hazard ratios (HRs) were noted among participants aged < 60 years, current smokers, and current alcohol consumers. Stratification by diabetes status revealed a markedly stronger increase in mortality risk among non-diabetic individuals, particularly in the FLI cohort (HR = 2.76, 95% CI: 2.00–3.80; P for interaction = 0.008). This suggests that ACR may act as an early warning signal of adverse outcomes even before the onset of overt diabetes.

For cardiovascular mortality, elevated ACR was also significantly associated with increased risk across all subgroups, with particularly strong associations observed in females, current smokers, and alcohol consumers; detailed results are presented in Supplementary Table S9. Stratification by FIB-4 score demonstrated the strongest association among individuals at high risk of liver fibrosis (F3–F4), with statistically significant interaction, indicating that the degree of liver fibrosis may modulate the ACR–CVD mortality relationship. Overall, although most subgroup interactions did not reach statistical significance, certain populations—including non-diabetic individuals, females, those with pre-existing CVD, and those at high risk of liver fibrosis–may be more sensitive to the detrimental effects of elevated ACR, providing potential clues for targeted interventions.

To verify the robustness of our findings, sensitivity analyses were performed, with detailed results presented in Supplementary Tables S11–S16 and Figure S3. These analyses included: recalculating ACR using published calibration equations to account for assay drifts across survey cycles (Table S11); excluding participants who died within the first two years of follow-up, excluding those with baseline malignancies (Table S12), and additionally adjusting for the use of renin–angiotensin–aldosterone system (RAAS) blockers and statins (Tables S13-S14). Notably, exclusion of deaths within the first two years did not materially alter the results, and additional adjustment for RAAS blockers and statins yielded consistent findings. We also re-evaluated the associations by categorising ACR into quartiles and visualised survival differences using Kaplan–Meier curves (Table S15 and Figure S3), which showed a clear dose–response trend. In addition, competing risk models were applied to cardiovascular mortality (Table S16), considering non-cardiovascular death as a competing event; results remained directionally consistent. Across all three MASLD-defined cohorts, the associations between elevated ACR (≥ 30 mg/g) and both all-cause and cardiovascular mortality remained statistically significant after adjustment for multiple covariates. Multicollinearity among covariates, including ACR, was evaluated, and results indicated that potential correlations did not materially affect the associations reported (Supplementary Table S10). Collectively, these complementary sensitivity analyses support the robustness and internal consistency of our main findings across different analytical scenarios.

ACR-related pathways in the associations of diabetes, hypertension, and mortality

After adjusting for potential confounders (Table 3), we explored whether ACR might act as a potential mediator in the associations of diabetes and hypertension with all-cause and cardiovascular mortality. The conceptual framework of these potential pathways is illustrated in the DAG (Figure S4). Across MASLD cohorts, the indirect effect of diabetes on all-cause mortality via ACR (average causal mediation effect, ACME) was statistically significant, suggesting that part of the observed association could be related to ACR. In the FLI and USFLI cohorts, the average direct effect (ADE) was not statistically significant, which is compatible with the notion that ACR may contribute to the association, although causality cannot be inferred. The proportion mediated ranged from 33.4% to 49.2% (all P < 0.01). For hypertension, the indirect effects via ACR were also statistically significant, with proportion mediated ranging from 18.9% to 24.9% (all P < 0.01), indicating potential partial mediation. For cardiovascular mortality, mediation analyses for diabetes showed significant indirect effects via ACR across MASLD cohorts, with proportion mediated ranging from 26.1% to 33.0% (all P ≤ 0.01). Some ADEs were not statistically significant, but the total associations remained consistent. Similarly, the indirect effects of hypertension on cardiovascular mortality via ACR were significant, with proportion mediated ranging from 12.1% to 19.0% (all P ≤ 0.01). Notably, sensitivity analyses (Supplementary Table S17) using medsens suggested that the observed ACME estimates could be influenced by unmeasured confounding, and therefore these results should be interpreted cautiously. Furthermore, the fully adjusted model, including additional clinical covariates such as eGFR, ALT, AST, and lipid parameters, yielded generally consistent results with the minimal adjustment model (Supplementary Table S18), supporting the robustness of the observed mediation patterns. Furthermore, we performed weighted mediation analyses to account for potential selection biases, generating results for both minimally adjusted (Table S19) and fully adjusted models (Table S20). Across these weighted analyses, the ACME, ADE, total effect, and proportion mediated were generally similar to the unweighted analyses, suggesting that ACR may play a role in the associations of diabetes and hypertension with mortality, with indirect effects remaining notable in most cohorts.

Table 3.

Mediation analysis of the role of ACR in the associations between diabetes or hypertension and mortality in the MASLD cohorts

Exposure	MASLD Cohort	Indirect Effect (ACME) (95% CI)	P value	Direct Effect (ADE) (95% CI)	P value	Total Effect (95% CI)	P value	Proportion Mediated (95% CI)	P value
A. All-cause Mortality
Diabetes	FLI	0.0191 (0.0127, 0.0300)	< 0.001	0.0197 (− 0.0014, 0.0400)	0.074	0.0388 (0.0183, 0.0600)	< 0.001	0.492 (0.274, 1.070)	< 0.001
	USFLI	0.0198 (0.0128, 0.0300)	< 0.001	0.0236 (− 0.0008, 0.0500)	0.056	0.0434 (0.0199, 0.0700)	0.002	0.457 (0.255, 1.030)	0.002
	HSI	0.0134 (0.0086, 0.0200)	< 0.001	0.0268 (0.0107, 0.0400)	< 0.001	0.0402 (0.0248, 0.0600)	< 0.001	0.334 (0.199, 0.590)	< 0.001
Hypertension	FLI	0.0113 (0.0074, 0.0200)	< 0.001	0.0340 (0.0132, 0.0500)	< 0.001	0.0452 (0.0241, 0.0600)	< 0.001	0.249 (0.147, 0.490)	< 0.001
	USFLI	0.0128 (0.0082, 0.0200)	< 0.001	0.0384 (0.0132, 0.0600)	0.004	0.0512 (0.0262, 0.0800)	< 0.001	0.249 (0.139, 0.510)	< 0.001
	HSI	0.0077 (0.0048, 0.0100)	< 0.001	0.0329 (0.0176, 0.0500)	< 0.001	0.0406 (0.0247, 0.0600)	< 0.001	0.189 (0.106, 0.330)	< 0.001
B. Cardiovascular Mortality
Diabetes	FLI	0.0073 (0.0042, 0.0100)	< 0.001	0.0152 (0.0018, 0.0300)	0.022	0.0225 (0.0093, 0.0400)	< 0.001	0.324 (0.172, 0.760)	< 0.001
	USFLI	0.0070 (0.0034, 0.0100)	< 0.001	0.0143 (− 0.0013, 0.0300)	0.064	0.0213 (0.0060, 0.0400)	0.012	0.330 (0.124, 1.020)	0.012
	HSI	0.0043 (0.0016, 0.0100)	0.002	0.0121 (0.0017, 0.0200)	0.022	0.0164 (0.0063, 0.0300)	0.002	0.261 (0.091, 0.720)	0.004
Hypertension	FLI	0.0041 (0.0022, 0.0100)	< 0.001	0.0175 (0.0041, 0.0300)	0.012	0.0216 (0.0085, 0.0300)	< 0.001	0.190 (0.087, 0.530)	< 0.001
	USFLI	0.0041 (0.0018, 0.0100)	0.002	0.0240 (0.0077, 0.0400)	0.004	0.0281 (0.0124, 0.0400)	< 0.001	0.146 (0.056, 0.380)	0.002
	HSI	0.0023 (0.0008, 0.0050)	0.004	0.0166 (0.0068, 0.0300)	< 0.001	0.0189 (0.0094, 0.0300)	< 0.001	0.121 (0.041, 0.270)	0.004

All mediation models were adjusted for potential confounders. For analyses with diabetes and hypertension as the exposure, adjusted covariates included: age, sex, race, poverty-income ratio (PIR), education, marital status, smoking status, and drinking status

Abbreviations: MASLD metabolic dysfunction-associated steatotic liver disease, ACR albumin-to-creatinine ratio, CI confidence interval, FLI Fatty Liver Index, USFLI United States Fatty Liver Index, HSI Hepatic Steatosis Index, ACME average causal mediation effect, ADE average direct effect

Overall, these analyses suggest a potential mediating role of ACR in the associations of diabetes and hypertension with mortality; however, causal inferences are limited due to the observational study design and potential residual confounding.

Model performance evaluation and key predictor identification for mortality risk in MASLD

We developed all-cause mortality risk prediction models using four machine learning algorithms—RSF, CoxBoost, GBM, and Lasso-Cox regression. All models demonstrated good predictive performance across MASLD cohorts, as detailed in Supplementary Table S21, with C-index values exceeding 0.83 and time-dependent AUCs above 0.81. RSF demonstrated the highest discrimination in both the FLI and USFLI cohorts. In the HSI-defined cohort, all models showed comparable AUC values. Variable importance analysis indicated that age was the most critical predictor of all-cause mortality across all models. Notably, ACR classification, reflecting proteinuria levels, consistently ranked highly, highlighting its central role in predicting mortality risk among MASLD patients (Figure S5). Additionally, kidney function indicators and inflammation-related markers were consistently among the top-ranked predictors across all models.

To address the limitation that machine learning models could not directly incorporate survey weights, we further validated the top 20 important predictors identified by machine learning models using survey-weighted Cox regression analyses, confirming the statistical significance of the majority of these variables (Tables S22–S24). Across FLI, USFLI, and HSI cohorts, variables such as ACR, age, and eGFR remained highly significant, supporting the robustness of the machine learning selected predictors in a nationally representative context.

To examine whether incorporating ACR could improve model performance, we calculated the integrated discrimination improvement (IDI) and net reclassification improvement (NRI) for the machine learning models with and without ACR (Table S21). Across the three MASLD cohorts, adding ACR did not lead to a statistically significant enhancement in discrimination or reclassification performance, which may reflect shared predictive information or potential collinearity between ACR and other kidney function indicators (e.g., eGFR, UREA, and CREA). To further assess model calibration and clinical utility, we constructed calibration plots (Figure S6) and performed DCA (Figure S7). The calibration assessment showed that the predicted survival probabilities were generally consistent with the observed outcomes. When the predicted survival probability was below approximately 50%, a slight overestimation of survival was observed, which is likely attributable to the limited number of events in the low-risk range and the estimation variability inherent in the resampling-based calibration process, rather than systematic model miscalibration. The DCA results indicated that the net clinical benefit of these models was modest, showing some benefit only when the threshold probability was below about 25%. Moreover, the 5-year DCA curves were consistently higher than the 10-year curves, suggesting a relatively better predictive utility for short-term risk estimation.

Discussion

In this nationally representative cohort of individuals with MASLD, we observed that elevated ACR levels were significantly associated with increased risks of both all-cause and cardiovascular mortality, with consistent associations across three commonly used MASLD definitions. These associations remained robust even after extensive adjustment for demographic characteristics, metabolic parameters, and clinical comorbidities. Notably, mediation analysis further suggested that ACR may partially account for the association of diabetes and hypertension with mortality, with nearly 50% of the diabetes–mortality association statistically attributed to ACR. Furthermore, stratified analyses demonstrated that mortality risks progressively increased across higher FIB-4 stages, reinforcing the prognostic value of hepatic fibrosis in MASLD. Overall, these findings indicate that ACR serves not only as a critical prognostic marker in MASLD populations at high risk but also may reflect a key biological pathway underlying adverse outcomes associated with metabolic dysfunction. Combined assessment of ACR and fibrosis indices may provide complementary insights into disease progression and prognosis, with implications for early risk stratification and targeted interventions in clinical practice.

In this study, we used three widely applied non-invasive indices—FLI, USFLI, and HSI—to identify individuals with MASLD. These indices offer good accessibility and reproducibility in large-scale population studies [22–34]. However, as surrogate markers based on metabolic and laboratory parameters, they are subject to a degree of misclassification compared with liver histology, the gold standard. Their sensitivity and specificity may vary, particularly in detecting mild steatosis or across populations with different ethnic or physiological characteristics [35, 36]. Biopsy-based evidence indicates that while FLI and HSI effectively discriminate the presence of steatosis, they are less accurate for quantifying changes in liver fat content and may be influenced by hepatic fibrosis or steatohepatitis [37]. At the population level, approximately 10–16% of individuals across the MASLD cohorts had elevated ACR (≥ 30 mg/g), indicating a notable proportion at increased cardiometabolic risk. Consequently, these non-invasive indices provide a feasible approach for large-scale epidemiological research, but their use for individual risk assessment in clinical settings should be interpreted with caution.

Although our study is based on NHANES, a U.S.-representative cohort, the generalizability of the findings to other populations requires careful consideration. The predictive performance of ACR and the three non-invasive hepatic steatosis indices (FLI, USFLI, and HSI) was robust across cohorts derived from different populations—FLI was developed in Italy (European), HSI in Korea (Asian), and USFLI in a multi-ethnic U.S. population—suggesting potential broad applicability. However, genetic predispositions, dietary habits, and healthcare systems vary across populations and could influence the observed associations. For example, gene–diet interactions modulate MASLD/NAFLD risk, with healthy plant-based diets reducing risk even in individuals with high genetic susceptibility [38]. Dietary patterns, such as meat- and flour-based diets versus prudent diets, significantly influence MASLD development in Asian populations [39]. Moreover, the PNPLA3 I148M (rs738409 C >G) risk allele (G) is prevalent worldwide, with frequency varying considerably across regions, influencing the risk of developing MASLD and other forms of steatotic liver disease [40]. Taken together, these observations suggest that while ACR is likely a useful prognostic marker in diverse populations, the strength of its associations with adverse outcomes may vary depending on population-specific genetic, dietary, and healthcare factors.

The kidney’s dense microvascular architecture renders it highly susceptible to microvascular injury, and glomerular endothelial cell dysfunction often precedes the onset of albuminuria [41]. Albuminuria reflects early endothelial dysfunction and is a recognised marker of microvascular damage, closely linked to the development of atherosclerosis and heart failure [42], indicating that the cardiovascular system is similarly affected [43]. Within the context of MASLD-related metabolic derangements, such damage may be further exacerbated. The pathogenesis of albuminuria is closely associated with mitochondrial oxidative stress within glomerular endothelial cells; inflammation induced by oxidative stress can damage podocytes, facilitating albumin leakage [44]. Moreover, podocytes may release endothelin-1, which activates endothelin A receptors on adjacent endothelial cells, thereby promoting mitochondrial dysfunction and contributing to a vicious cycle of podocyte apoptosis and glomerulosclerosis [45]. Likewise, MASLD progression is tightly linked to hepatic mitochondrial oxidative stress, with mechanisms involving oxidative damage and redox imbalance leading to apoptosis and inflammation [46, 47]. Taken together, these parallel mechanisms suggest a hypothesis that albuminuria in MASLD may be mechanistically connected to concomitant mitochondrial dysfunction along a liver–kidney axis, wherein mutually reinforcing oxidative stress and inflammation could contribute to a “hepatorenal injury” phenotype. Additionally, albuminuria is strongly associated with insulin resistance and MetS [48]—core mechanisms underpinning MASLD and its complications [49]. Notably, albuminuria may not only be a consequence of insulin resistance, but also exacerbate it, perpetuating a cycle of metabolic toxicity and increased mortality risk in MASLD populations.

Our study further demonstrated that ACR accounted for a substantial proportion of the observed associations between diabetes, hypertension, and mortality, explained up to 49% of the diabetes–mortality association. Previous research has shown that albuminuria is not only a consequence of insulin resistance, but also a shared risk factor for MetS, diabetes, and CVD [50]. Components of the MetS—such as obesity, impaired fasting glucose, hypertension, and hypertriglyceridaemia—are strongly associated with increased risk of albuminuria [51], supporting the notion that it may reflect a common pathway of metabolic dysfunction. Importantly, elevated ACR may interact with metabolic dysfunction to further aggravate inflammation and oxidative stress, potentially creating a self-reinforcing feedback loop that accelerates disease progression. Given that ACR testing is simple and cost-effective, its potential utility in early warning, risk stratification, and intervention for metabolic diseases deserves greater clinical attention. Nonetheless, as this was an observational study, the mediation results should be interpreted with caution. Sensitivity analyses using the medsens function indicated that the observed ACME estimates could be influenced by residual confounding, meaning that the proportion mediated may be over- or underestimated. Residual confounding, particularly in the relationship between ACR and mortality, cannot be fully excluded, and the key assumptions underlying mediation analyses—such as the absence of unmeasured mediator–outcome confounding—should be acknowledged. While a previous study using NHANES data reported the association between ACR and mortality in MASLD [52], it did not examine the extent to which ACR explained diabetes- and hypertension-related mortality associations. By integrating mediation analyses across multiple MASLD definitions, our study provides a more comprehensive understanding of ACR’s prognostic significance in this context.

Across the three MASLD cohorts, subgroup analyses demonstrated that elevated ACR (≥ 30 mg/g) was consistently associated with higher all-cause and cardiovascular mortality risks, with particularly pronounced effects observed in individuals under 60 years of age, females, current smokers, and current drinkers. Notably, female patients with MASLD and concomitant renal impairment appeared to be at greater cardiovascular risk, suggesting a higher vulnerability compared with males. However, a cohort study by Toth-Manikowski et al. reported lower all-cause and cardiovascular mortality among women than men with CKD [53], which contrasts with our findings. This discrepancy may reflect diminished oestrogen-mediated vascular protection in women with coexisting MASLD and CKD. Although elevated ACR remained associated with increased mortality risk among diabetic participants, the strength of this association was somewhat attenuated compared to non-diabetics. This suggests that ACR may serve as a more sensitive indicator of cardiorenal risk in metabolically less overt populations and a potential early warning marker before the clinical manifestation of diabetes. In our stratified analysis by FIB-4 categories, both all-cause and cardiovascular mortality risks increased with advancing fibrosis stages, especially in the high-risk (F3–F4) group across all MASLD definitions. This trend highlights the prognostic importance of liver fibrosis in MASLD. Cardiovascular mortality was not estimable in the low-risk (F0–F2) group, likely due to few events, underscoring the protective role of early-stage disease. Given that ACR reflects systemic inflammation and metabolic stress, it may serve as a non-invasive marker to help identify individuals at higher risk of hepatic progression. Future studies could investigate combining ACR with fibrosis indices to enhance risk stratification and guide advanced diagnostic assessments. Sensitivity analyses further confirmed the robustness of our findings: the association between elevated ACR and increased all-cause and cardiovascular mortality remained significant even after excluding early deaths and participants with cancer, supporting its potential role in identifying high-risk individuals with MASLD. Collectively, these findings suggest that ACR could be incorporated into clinical practice as a pragmatic risk stratification tool. For instance, an ACR threshold of ≥ 30 mg/g may help identify MASLD patients at elevated risk, prompting more frequent monitoring and earlier interventions, such as intensified metabolic management, lifestyle modifications, and closer cardiovascular and renal follow-up. Furthermore, integrating ACR with established fibrosis indices, such as FIB-4, may enhance risk stratification and support clinicians in tailoring diagnostic and therapeutic strategies to individual patient profiles. Future prospective studies are warranted to establish evidence-based ACR thresholds and to evaluate whether ACR-guided management can improve clinical outcomes in MASLD patients. Overall, ACR demonstrated greater predictive utility in certain clinical subgroups, highlighting the importance of dynamic monitoring in MASLD management to enable early detection of latent cardiovascular and renal risks, thereby improving patient outcomes.

The use of machine learning models, including RSF, CoxBoost, GBM, and Lasso-Cox, offers advantages over traditional Cox regression by capturing non-linear relationships, handling high-dimensional data, and providing variable importance rankings. Most survival models demonstrated good predictive performance across the three MASLD definitions, with consistent C-index, AUC, and Brier score metrics, suggesting that ACR, particularly when combined with other clinical variables, may contribute meaningfully to mortality risk prediction. Notably, in contexts where comprehensive metabolic profiling may be impractical, ACR alone showed potential for identifying individuals at higher risk, indicating a possible role in initial risk stratification. However, the observed improvements in IDI and NRI were limited, and calibration and decision curve analyses indicated only modest clinical benefit, underscoring the need for cautious interpretation and external validation in future work. Future studies may focus on developing risk scores centred on ACR or exploring its integration into existing risk assessment frameworks to further evaluate its predictive utility in MASLD populations.

This study has several strengths. First, it was based on nationally representative NHANES data and employed statistical weighting to account for the complex sampling structure. Mortality outcomes were derived from a reliable and authoritative death registry, ensuring the validity and generalisability of the results. Second, the use of three widely recognised MASLD definitions and the consistency of findings across cohorts strengthen the robustness of the conclusions. However, several limitations should be acknowledged. Although the use of three non-invasive indices enhanced the robustness of our findings, these surrogate markers are not diagnostic gold standards and lack imaging or histological validation, potentially introducing misclassification bias. Indeed, only modest agreement was observed among the FLI, USFLI, and HSI classifications, underscoring the possibility of index-derived misclassification. Moreover, because these indices incorporate metabolic and liver-related parameters (e.g., BMI, triglycerides, ALT/AST), which are themselves associated with both ACR and mortality, there remains a possibility of circularity and collinearity. ACR was assessed only once, and transient physiological or behavioural factors may have influenced the measurements. Moreover, a substantial number of participants were excluded due to missing data. To provide context, we characterised these excluded individuals and compared them with the analytic sample, thereby illustrating the extent to which selection bias may be present. Importantly, part of the missingness in NHANES stems from its survey design, as not all laboratory tests were performed in every cycle. This design-related missingness inevitably reduces sample size and may introduce bias, representing an inherent limitation of secondary analyses based on NHANES. Finally, despite extensive adjustment for confounders, the possibility of residual confounding cannot be entirely excluded. Future studies with longitudinal follow-up and gold-standard diagnostic tools are needed to further validate these findings.

Conclusion

This study systematically evaluated the prognostic utility of ACR in predicting mortality among individuals with MASLD. Elevated ACR was identified as an independent predictor of all-cause and cardiovascular mortality and may partly contribute to the observed associations of diabetes and hypertension with adverse outcomes. Machine learning analyses consistently supported age and ACR as key predictors of all-cause mortality. Additionally, mortality risks increased progressively with higher FIB-4 stages, emphasizing the prognostic relevance of hepatic fibrosis. These results support the complementary use of ACR and fibrosis indices for risk stratification. Importantly, ACR testing is simple, inexpensive, and widely available, making its implementation in primary care and hepatology clinics a feasible and pragmatic strategy. Future studies should validate these findings and explore integrated risk models combining ACR and fibrosis markers for early detection and precision management in MASLD populations.

Abbreviations

MASLD

Metabolic dysfunction-associated steatotic liver disease

NAFLD

Non-alcoholic fatty liver disease

ACR

Albumin-to-creatinine ratio

NHANES

National Health and Nutrition Examination Survey

FLI

Fatty Liver Index

USFLI

United States Fatty Liver Index

HSI

Hepatic Steatosis Index

CVD

Cardiovascular disease

CKD

Chronic kidney disease

NCHS

National Center for Health Statistics

KDIGO

Kidney Disease: Improving Global Outcomes

ICD-10

International Classification of Diseases, 10th Revision

MetS

Metabolic syndrome

SBP

Systolic blood pressure

DBP

Diastolic blood pressure

NCEP ATP III

National Cholesterol Education Program Adult Treatment Panel III

BMI

Body mass index

WC

Waist circumference

HbA1c

Glycosylated hemoglobin type A1c

TC

Total cholesterol

TG

Triglyceride

LDL-C

Low-density lipoprotein cholesterol

AST

Aspartate aminotransferase

ALT

Alanine aminotransferase

GGT

γ-glutamyl transpeptidase

UREA

Serum urea

CREA

Creatinine

UA

Uric acid

eGFR

Estimated glomerular filtration rate

FIB-4

Fibrosis-4 index

NLR

Neutrophil-to-lymphocyte ratio

SII

Systemic immune-inflammation index

SD

Standard deviations

IQR

Interquartile ranges

Log(ACR)

Natural logarithm of the albumin-to-creatinine ratio

RSF

Random survival forest

VIMP

Variable importance

GBM

Gradient boosting machine

Lasso

Least absolute shrinkage and selection operator

DAG

Directed acyclic graph

VIF

Variance Inflation Factor

RAAS

Renin–angiotensin–aldosterone system

CI

Confidence interval

HR

Hazard ratio

ACME

Average causal mediation effect

ADE

Average direct effect

RSF

Random survival forest

AUC

Area under the curve

DCA

Decision curve analyses

IDI

Integrated discrimination improvement

NRI

Net reclassification improvement

GBM

Gradient boosting machine

C-index

Concordance index

FBG

Fasting blood glucose

SIRI

Systemic inflammation response index

RAAS

Renin–angiotensin–aldosterone system

Authors’ contributions

ZHJ, LM and OC conceived the study; ZHJ and HH drafted and revised the manuscript; LM conceptualized the article; ZHL and PNQ analyzed the data; HMF participated in data acquisition. The final version of the manuscript has been read and approved by all authors.

Funding

This study was supported in part by grants from the Key R & D program Natural Science Foundation of Guangxi Province under Grant No. AB19110007.

Data availability

The datasets used and evaluated in this study can be obtained from the corresponding author upon reasonable request. The analysis codes and variable mappings used to process and analyse NHANES data are also available upon reasonable request to ensure reproducibility.

Declarations

Ethics approval and consent to participate

All survey protocols were approved by the National Center for Health Statistics Ethics Review Board. All participants provided written informed consent before participation.

Competing interests

The authors declare no competing interests.