A cardiometabolic risk phenotype index integrating insulin resistance, systemic inflammation, and psychological burden for mortality risk stratification in metabolic dysfunction-associated steatotic liver disease: an exploratory longitudinal analysis of a nationally representative cohort

Abstract

Background

Metabolic dysfunction-associated steatotic liver disease (MASLD) is being increasingly recognized as a systemic cardiometabolic disorder rather than a liver-limited condition. However, the combined association of insulin resistance, systemic inflammation, and depressive symptoms with mortality remains unclear. Therefore, we developed an insulin resistance-systemic inflammation-psychological burden (ISP) index to evaluate its associations with all-cause mortality (ACM) and cardiovascular mortality (CVM).

Methods

We analyzed data from 3,187 adults with MASLD in the 2005–2018 National Health and Nutrition Examination Survey. A random forest model based on mean decrease in accuracy was used to select representative markers from six insulin resistance surrogates and five systemic inflammation indices, which were combined with the PHQ-9 score to construct the ISP index. Weighted Kaplan–Meier and Cox analyses, restricted cubic splines, subgroup and sensitivity analyses, and interpretable machine learning with 5-fold cross-validation were performed.

Results

During a median 8.25-year follow-up, 382 deaths occurred, including 114 cardiovascular deaths. The metabolic score for insulin resistance (METS-IR) and systemic inflammation response index (SIRI) were selected and combined with the PHQ-9 score to construct the ISP index. In fully adjusted Cox models, the highest ISP tertile was associated with higher risks of both ACM (HR = 1.63, 95% CI: 1.12–2.38) and CVM (HR = 2.88, 95% CI: 1.36‒6.10) compared with the lowest tertile. Eight ISP-based machine-learning models were internally evaluated using 5-fold cross-validation, with the GBM model yielding 10-year AUCs of 0.869 for ACM and 0.883 for CVM. SHAP analysis of the GBM model revealed that the ISP index ranked second in feature importance after age.

Conclusions

The ISP index was positively associated with the risks of ACM and CVM among individuals with MASLD. However, because the ISP index was constructed and internally evaluated within the same sample, its apparent predictive performance may be optimistic owing to overfitting. Therefore, the ISP index requires external validation in larger, independent prospective cohorts.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12944-026-03048-4.

Introduction

Metabolic dysfunction-associated steatotic liver disease (MASLD) has replaced nonalcoholic fatty liver disease as the preferred nomenclature for steatotic liver disease driven by metabolic dysfunction. The updated terminology aligns disease naming with underlying metabolic biology and avoids stigmatizing language [1]. This transition is clinically important because fatty liver disease is highly prevalent, affecting approximately one-third of adults globally, and its prevalence has increased over time [2]. The MASLD framework therefore describes not only a liver phenotype but also a common cardiometabolic condition associated with obesity, dysglycemia, hypertension, atherogenic dyslipidemia, and systemic metabolic dysfunction.

The clinical relevance of MASLD is largely determined by its extrahepatic consequences. MASLD is increasingly recognized as a systemic metabolic disorder with cardiovascular and malignant complications, and fatty liver disease requires a multidisciplinary perspective because its consequences extend beyond hepatology [3, 4]. This perspective is especially relevant for cardiovascular risk assessment. Fatty liver disease is associated with an increased long-term risk of fatal and nonfatal cardiovascular events, and cardiovascular disease is a principal cause of death in this population [5, 6]. Accumulating evidence has demonstrated that insulin resistance, systemic inflammation, and psychological burden are associated with adverse outcomes among patients with MASLD [7–12]. Given the heterogeneous and systemic nature of MASLD, these single-domain markers may not sufficiently capture the multidimensional risk burden that contributes to long-term adverse outcomes. Therefore, we sought to develop a novel insulin resistance–systemic inflammation–psychological burden (ISP) index to comprehensively reflect these interconnected pathophysiological domains in patients with MASLD. As an integrated multidimensional measure, the ISP index may have potential for identifying higher-risk phenotypes and supporting exploratory long-term risk stratification.

The rationale for the ISP index rests on the possibility that insulin resistance, inflammation, and depressive symptoms do not act simply as parallel risk markers but converge on shared cardiometabolic pathways. Insulin resistance promotes ectopic lipid accumulation, atherogenic dyslipidemia, adipose tissue dysfunction, endothelial stress, and chronic low-grade inflammation, whereas systemic inflammatory activation may translate metabolic stress into vascular injury [13–15]. Psychological burden may further amplify this risk state through neuroendocrine activation, autonomic imbalance, inflammatory signaling, sleep disruption, physical inactivity, dietary patterns, and adherence-related pathways [16, 17]. The central premise of this study is that the joint burden of these domains may identify a higher-risk MASLD phenotype than any single domain considered alone. By combining these domains, the ISP index may identify MASLD patients in whom metabolic dysfunction is accompanied by inflammatory activation and psychological burden, a pattern that appears particularly relevant to mortality risk among MASLD patients. Although the METS-IR, SIRI, and depressive symptom burden have each been examined separately in relation to cardiometabolic outcomes, MASLD-related prognosis is unlikely to be determined by a single risk domain alone. The updated MASLD nomenclature emphasizes metabolic dysfunction as the defining biological context of steatotic liver disease, and recent clinical guidance further frames MASLD as a multisystem condition requiring cardiometabolic risk assessment rather than liver-focused evaluation alone [1, 18]. In parallel, the cardiovascular–kidney–metabolic framework highlights the interdependence of metabolic dysfunction, systemic vascular injury, kidney-related risk, and cardiovascular outcomes [19]. These conceptual developments support the need for integrated phenotyping approaches that can capture coexisting metabolic, inflammatory, and neurobehavioral vulnerability. Therefore, the ISP index was designed not to replace the METS-IR, SIRI, or PHQ-9 score but to summarize their joint burden into a single biologically interpretable phenotype that may better reflect the systemic nature of MASLD.

In this context, we developed the ISP index by integrating multiple indicators of insulin resistance, systemic inflammation, and psychological burden to evaluate its associations with mortality and explore its potential utility for mortality risk stratification among individuals with MASLD.

Methods

Study population

This study used data from the National Health and Nutrition Examination Survey (NHANES), a continuous, nationally representative survey conducted by the National Center for Health Statistics of the Centers for Disease Control and Prevention. The NHANES uses a complex, multistage probability sampling design to assess the health and nutritional status of the civilian noninstitutionalized U.S. population. In this study, data from the NHANES 2005–2018 cycles were used. After excluding participants who were younger than 18 years, those lacking sufficient information to define MASLD, those with viral hepatitis or excessive alcohol consumption, and those with missing covariate or mortality linkage information, the final analytical cohort comprised 3,187 adults with MASLD. Detailed participant selection is shown in Fig. S1. The present analysis used publicly available deidentified data and followed NHANES analytic guidance for complex survey data.

Definition of MASLD

MASLD was defined according to the multisociety Delphi consensus statement [1]. Specifically, the diagnosis of MASLD required the presence of hepatic steatosis in combination with at least one cardiometabolic risk factor after participants with viral hepatitis or excessive alcohol consumption were excluded. In large-scale epidemiological studies, defining hepatic steatosis on the basis of imaging or histopathological assessment is often not feasible. Therefore, the fatty liver index (FLI), a convenient and validated noninvasive surrogate marker calculated using γ-glutamyl transferase (GGT), triglyceride (TG), waist circumference (WC), and body mass index (BMI), was used to assess hepatic steatosis, which was defined as an FLI ≥ 60. The validity of the FLI has been confirmed in multiple studies [20–22], and it has been recommended by the EASL-EASD-EASO clinical practice guidelines as an acceptable alternative method for steatosis assessment [23]. The FLI was calculated using the following formula:

Adult cardiometabolic risk factors were defined as the presence of at least one of the following conditions: (a) overweight/obesity (BMI ≥ 25 kg/m²) or central obesity (WC > 94 cm for men and > 80 cm for women); (b) dysglycemia, including fasting blood glucose ≥ 5.6 mmol/L, 2-h post-load glucose ≥ 7.8 mmol/L, HbA1c ≥ 5.7%, or a previous diagnosis of diabetes; (c) hypertension (blood pressure ≥ 130/85 mmHg or use of antihypertensive medication); (d) hypertriglyceridemia (plasma triglycerides ≥ 1.70 mmol/L or use of lipid-lowering therapy); or (e) reduced high-density lipoprotein cholesterol (HDL-C) levels (≤ 1.0 mmol/L for men and ≤ 1.3 mmol/L for women) or use of lipid-lowering therapy [1].

Selection of representative insulin resistance surrogates and systemic inflammation indices

To identify representative insulin resistance and systemic inflammation markers on the basis of their predictive contributions, we compared six insulin resistance surrogate markers and five systemic inflammation indices. The insulin resistance surrogate markers included the triglyceride-glucose (TyG) index, estimated glucose disposal rate (eGDR), metabolic score for insulin resistance (METS-IR), homeostasis model assessment of insulin resistance (HOMA-IR), cholesterol, high-density lipoprotein, and glucose (CHG) index, and lipid accumulation product (LAP). In addition, the systemic inflammation indices included the neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), monocyte-to-lymphocyte ratio (MLR), systemic immune-inflammation index (SII), and systemic inflammation response index (SIRI). The detailed calculation formulas for these indicators are presented in Fig. S2 [24–28].

The random forest mean decrease in accuracy (MDA) was used to estimate the relative contribution of each candidate marker to predicting ACM and CVM. MDA quantifies the reduction in prediction accuracy after permuting a given variable; a larger decrease indicates a greater contribution to model performance. When the highest-ranking markers differed between ACM and CVM, their outcome-specific MDA rankings and absolute MDA differences were jointly considered to identify the marker with the most balanced predictive contribution across both outcomes. On this basis, METS-IR and SIRI were selected as representative markers of insulin resistance and systemic inflammation, respectively, to construct the ISP index (Fig. 1).

Fig. 1.

Predictive contribution ranking of insulin resistance surrogates and systemic inflammation markers on the basis of the mean decrease in accuracy. Abbreviations: A Insulin resistance surrogates for ACM; B insulin resistancesurrogates for CVM; C systemic inflammatory markers for ACM; and D systemic inflammatory markers for CVM, all in patients with MASLD, metabolic dysfunction-associated steatotic liver disease; TyG, triglyceride-glucose index; CHG, cholesterol, high-density lipoprotein, and glucose; LAP, lipid accumulation product; eGDR, estimated glucose disposal rate; METS-IR, metabolic score for insulin resistance; HOMA-IR, homeostasis model assessment of insulin resistance; NLR, neutrophil-to-lymphocyte ratio; PLR, platelet-to-lymphocyte ratio; MLR, monocyte-to-lymphocyte ratio; SII, systemic immune-inflammation index; SIRI, systemic inflammation response index

This selection is consistent with prior evidence supporting METS-IR as a surrogate marker of insulin resistance and visceral adiposity [29] and SIRI as a systemic inflammation marker associated with cardiovascular outcomes and mortality in individuals with MASLD [10, 30].

Assessment of psychological burden

Psychological burden was assessed using the PHQ-9 score. The PHQ-9 score evaluates the frequency of depressive symptoms across nine items, each scored from 0 to 3, yielding a total score ranging from 0 to 27. Higher scores indicate greater depressive symptom burden. PHQ-9 total scores were calculated only for participants with complete responses to all the PHQ-9 items [31]. In the present study, the PHQ-9 was treated as a continuous indicator of psychological burden rather than as a diagnostic psychiatric label. This terminology is important because the ISP index is intended to capture a cardiometabolic risk domain related to depressive symptoms, not to define a psychiatric disorder. The inclusion of the PHQ-9 was supported by evidence linking depressive symptoms with cardiovascular mortality in cardiometabolic populations [32, 33] and by evidence that depression, anxiety, or stress frequently coexist with fatty liver disease [34].

ISP index construction

To quantify the combined effect of insulin resistance, systemic inflammation, and psychological burden on mortality risk, multivariable Cox regression models were constructed separately for ACM and CVM. METS-IR, the SIRI, and the PHQ-9 score were included in the models. To eliminate the influence of different measurement units and scales across these indicators, all three variables were standardized as Z scores before being entered into the models. The Cox regression model is expressed as follows:

where h(t, X) represents the hazard function given the covariates X; h₀(t) represents the baseline hazard function; and β₁, β₂, and β₃ represent the regression coefficients of the METS-IR, SIRI, and PHQ-9 scores, respectively.

The regression coefficients of the METS-IR, SIRI, and PHQ-9 scores were then extracted from the ACM and CVM models. Within each outcome-specific model, these coefficients were normalized to obtain the relative weight of each indicator for the corresponding outcome [35, 36]. Specifically, the normalized weights in the ACM model were calculated as follows:

The normalized weights in the CVM model were calculated as follows:

The normalized weights of the same indicator from the ACM and CVM were then averaged to generate the final weights:

where , , and represent the final weights of the METS-IR, SIRI, and PHQ-9 scores in the composite index, respectively (weights are shown to two decimal places for presentation only). The insulin resistance–systemic inflammation–psychological burden (ISP) index is presented as follows:

This index, with component weights derived from the current NHANES analytical sample, was designed to comprehensively reflect the combined burden of insulin resistance, systemic inflammation, and psychological burden in relation to mortality risk.

Mortality outcomes

Mortality data were obtained from the NCHS public-use linked mortality files, which link NHANES participants to death certificate records from the National Death Index, with follow-up through December 31, 2019. Causes of death were classified according to the International Classification of Diseases, 10th Revision. ACM was defined as death from any cause, and CVM was defined as death attributable to heart disease.

Covariates

On the basis of the initial study design and evidence from the literature, the study covariates included demographic characteristics, lifestyle factors, socioeconomic status, medical history, and laboratory measurements. All variables were obtained from the Demographics, Examination, Laboratory, and Questionnaire datasets of the NHANES. Specifically, the demographic and socioeconomic variables included age, sex, race, educational level, marital status, and the poverty income ratio (PIR). Lifestyle-related variables included smoking status, alcohol consumption, and body mass index (BMI). Laboratory variables included fasting blood glucose (FBG), HbA1c, triglyceride (TG), total cholesterol (TC), low-density lipoprotein cholesterol (LDL-C), and high-density lipoprotein cholesterol (HDL-C), along with other variables used for calculating the six insulin resistance surrogate markers and five systemic inflammation indices. In addition, medical history variables included hypertension, diabetes mellitus (DM), and cancer, with detailed diagnostic criteria and definitions provided in Table S1. Fasting laboratory indicators were measured using standardized NHANES laboratory protocols. Blood samples were collected from participants after an overnight fast of 8–24 h and processed according to established procedures. Serum lipid profiles were measured using enzymatic methods on Roche/Hitachi analyzers, including the Roche Hitachi 717/912, Roche Modular P, and Roche/Hitachi Cobas 6000 systems, depending on the survey cycle. All laboratory assays were performed in certified laboratories following standardized quality assurance procedures. In addition, questionnaire-based data were collected using standardized interviewer training, structured questionnaires, and a computer-assisted personal interviewing (CAPI) system with built-in consistency checks to improve data reliability [37].

Statistical analyses

Survey-weighted analyses accounted for NHANES sampling weights, clustering, and stratification according to analytic recommendations. Continuous variables are summarized as medians with interquartile ranges, and categorical variables are summarized as weighted percentages. Differences across ISP tertiles were evaluated using the Kruskal‒Wallis test for continuous variables and Rao–Scott adjusted chi-square testsfor categorical variables. Statistical analyses were performed using R software. A two-sided P value < 0.05 was considered to indicate statistical significance.

Association of the ISP index with mortality risk

Weighted Kaplan‒Meier curves were used to compare survival probabilities across ISP tertiles, and log-rank tests were used to evaluate between-group differences. Weighted Cox proportional hazards models were then used to estimate hazard ratios and 95% confidence intervals for the associations of the ISP index with ACM and CVM. Model 1 was unadjusted. Model 2 was adjusted for age and sex. Model 3 was further adjusted for race, PIR, marital status, educational level, BMI, smoking status, alcohol consumption, hypertension, low-density lipoprotein cholesterol, diabetes mellitus, and cancer history. The ISP index was analyzed both as a continuous Z score and according to tertiles. To reduce potential reverse causality, sensitivity analyses were performed after excluding participants who died within the first year of follow-up. In addition, we examined the robustness of the associations between the ISP index and mortality outcomes using weighted propensity score matching and unweighted multivariable Cox regression, thereby assessing whether the main findings were sensitive to covariate balance or survey weighting. Stratified and interaction analyses were further performed to examine whether these associations varied across subgroups defined by sex, age, and diabetes status.

Dose‒response and subgroup analyses

Restricted cubic spline (RCS) analyses were used to examine potential nonlinear associations between the ISP index and mortality outcomes. These analyses were performed in the overall population and separately in subgroups defined by sex, age, and diabetes status.

Incremental predictive value analysis

Cox proportional hazards models were constructed to evaluate the incremental predictive value of the ISP index for ACM and the CVM. A base model including predefined clinical covariates was first developed, and an extended model was then constructed by adding the ISP index to the base model. Using the same 5-fold cross-validation framework and fold assignment as those used in subsequent machine-learning model development, out-of-fold risk scores were obtained from both models. The numbers of ACM and CVM events in each fold are provided in Table S2. The risk scores from each model were divided into tertiles to classify participants into low-, intermediate-, and high-risk groups. These tertile-based categories were used as research-oriented risk strata for model comparison and visualization, not as clinically established decision thresholds. Time-dependent continuous net reclassification improvement (NRI) with 95% confidence intervals and integrated discrimination improvement (IDI) at 10 years were used to assess whether adding the ISP index improved risk reclassification and discrimination beyond the base model. The continuous NRI was interpreted as a measure of net directional improvement in predicted risks without relying on predefined risk categories, whereas the IDI was interpreted as a measure of average improvement in discrimination. Therefore, statistically significant NRI or IDI values were not considered sufficient evidence for clinical implementation without external validation, calibration assessment, decision curve evaluation, or clinical impact testing.

Machine learning model development and interpretation

Eight prediction models were developed using the ISP index and clinically accessible variables: Cox regression, random survival forest (RSF), gradient boosting machine (GBM), survival support vector machine (SurvivalSVM), XGBoost, supervised principal components (SuperPC), partial least squares regression for Cox models (PLSR-Cox), and CoxBoost. The Cox model was used as the benchmark model.

For each machine learning model, the hyperparameters were tuned using a grid search within a 5-fold cross-validation framework. The mean C-index was calculated for each candidate hyperparameter combination, and the final hyperparameter setting for each algorithm was selected on the basis of the highest mean cross-validated C-index. To improve comparability across models, the same fold assignment was used for all the candidate models. In each round of cross-validation, continuous variables were standardized using parameters estimated from the training subset only. Specifically, the mean and standard deviation were calculated exclusively from the training subset and then applied to both the training subset and the corresponding validation subset, thereby minimizing information leakage during preprocessing.

Model reporting was guided by contemporary recommendations for transparent reporting and risk-of-bias assessment in prediction models using regression or machine learning [38, 39]. Model performance was evaluated using pooled out-of-fold predictions from the same 5-fold cross-validation framework, including the C-index, time-dependent 10-year area under the receiver operating characteristic curve, Brier score, calibration curves, and decision curve analysis.

SHapley Additive exPlanations (SHAP) values were used to interpret the selected GBM model. SHAP is a unified additive feature attribution framework that assigns each predictor a contribution to model output. In this study, SHAP summary plots, individual explanation plots, and ISP index dependence plots were generated to assess the global and individual-level contributions of the ISP index to ACM and CVM prediction. Finally, an exploratory web-based application was developed based on the fitted GBM models using the Shiny package in R to generate and visualize individualized model-estimated 10-year risks of ACM and CVM among individuals with MASLD.

Results

Characteristics of the participants

A total of 3,187 adults with MASLD were included in the analysis cohort. During a median follow-up of 8.25 years, 382 deaths occurred, corresponding to a weighted mortality proportion of 8.63%; among these deaths, 114 were attributable to cardiovascular causes. Participants were categorized into low, middle, and high ISP tertiles according to baseline ISP values (Table 1).

Table 1.

Characteristic of participants according to ISP index

Variables	Total (N = 3187)	Low Tertile 1 (−1.12, −0.32)	Middle Tertile 2 (−0.32, 0.13)	High Tertile 3 (> 0.13)	P-value
Age (years)	54.00(41.00,65.00)	54.00(43.00,64.00)	53.00(39.00,63.00)	55.00(42.00,66.00)	0.15
Sex, n(%)					0.01
Female	1496(45.14)	465(39.26)	520(47.50)	511(48.20)
Male	1691(54.86)	598(60.74)	542(52.50)	551(51.80)
BMI (kg/m2)	33.40(30.33,37.59)	30.73(28.80,33.06)	33.66(30.90,36.75)	37.55(32.88,42.90)	< 0.0001
Race, n(%)					< 0.001
Non-Hispanic Black	662(10.55)	282(13.50)	206(9.95)	174(8.42)
Mexican American	495(7.29)	188(8.56)	170(7.44)	137(5.94)
Non-Hispanic White	1522(72.15)	399(66.94)	515(72.06)	608(77.09)
Other Race	508(10.01)	194(11.00)	171(10.55)	143(8.54)
Education levels, n(%)					0.02
< high school	813(16.78)	249(14.48)	280(16.86)	284(18.85)
= high school	750(24.56)	230(21.61)	256(24.95)	264(26.90)
> high school	1624(58.66)	584(63.91)	526(58.19)	514(54.26)
PIR, n(%)					< 0.0001
≤ 1	617(12.49)	159(9.59)	202(12.35)	256(15.34)
1–3	1397(37.85)	422(32.32)	483(39.07)	492(41.75)
> 3	1173(49.66)	482(58.09)	377(48.58)	314(42.91)
Marital status, n(%)					< 0.0001
Never married	337(10.74)	95(8.04)	120(12.36)	122(11.60)
Married/living with partner	2095(70.11)	751(76.87)	700(70.36)	644(63.56)
Divorced/separated/widowed	755(19.15)	217(15.09)	242(17.28)	296(24.84)
Smoking status, n(%)					< 0.001
Never	1822(58.06)	668(62.22)	626(59.66)	528(52.54)
Former	964(30.43)	312(30.76)	308(28.57)	344(32.02)
Current	401(11.51)	83(7.02)	128(11.77)	190(15.43)
Alcohol consumption, n(%)					0.003
Never	583(15.07)	202(14.67)	202(15.00)	179(15.52)
Former	1023(28.72)	305(24.08)	336(28.20)	382(33.58)
Mild	1581(56.21)	556(61.25)	524(56.80)	501(50.90)
Hypertension, n(%)					< 0.0001
No	1305(44.34)	512(52.71)	460(47.23)	333(33.60)
Yes	1882(55.66)	551(47.29)	602(52.77)	729(66.40)
DM, n(%)					< 0.0001
No	2061(71.19)	773(81.29)	718(74.58)	570(58.30)
Yes	1126(28.81)	290(18.71)	344(25.42)	492(41.70)
Depressive symptoms, n(%)					< 0.0001
No	2889(91.64)	1062(99.91)	1002(95.34)	825(80.15)
Yes	298(8.36)	1(0.09)	60(4.66)	237(19.85)
Cancers, n(%)					0.01
No	2807(88.62)	965(91.58)	933(88.69)	909(85.77)
Yes	380(11.38)	98(8.42)	129(11.31)	153(14.23)
Laboratory data
FBG (mmol/L)	5.83(5.44,6.55)	5.66(5.33,6.11)	5.83(5.38,6.44)	6.16(5.61,7.27)	< 0.0001
HbA1c(%)	5.60(5.30,6.10)	5.50(5.30,5.90)	5.60(5.30,6.00)	5.80(5.40,6.50)	< 0.0001
TG (mmol/L)	1.55(1.12,2.17)	1.45(1.07,2.01)	1.57(1.14,2.17)	1.65(1.14,2.28)	0.004
TC (mmol/L)	4.97(4.29,5.69)	5.15(4.47,5.84)	4.99(4.29,5.74)	4.78(4.09,5.40)	< 0.0001
LDL-C (mmol/L)	2.97(2.38,3.62)	3.13(2.53,3.70)	3.03(2.38,3.67)	2.82(2.20,3.44)	< 0.0001
HDL-C (mmol/L)	1.19(1.01,1.40)	1.27(1.09,1.47)	1.16(1.01,1.37)	1.11(0.96,1.32)	< 0.0001
FLI	85.01(73.29,94.64)	74.74(67.30,82.87)	87.17(76.08,93.42)	94.47(84.23,98.75)	< 0.0001
Insulin Resistance Surrogates
TyG	8.94(8.57,9.31)	8.81(8.48,9.16)	8.94(8.57,9.30)	9.06(8.62,9.43)	< 0.0001
CHG	5.43(5.21,5.66)	5.35(5.16,5.56)	5.44(5.21,5.65)	5.50(5.28,5.77)	< 0.0001
LAP	76.96(55.45,108.08)	62.32(48.37,82.77)	78.42(58.26,104.65)	94.66(66.36,132.37)	< 0.0001
eGDR	5.45(4.04,8.02)	7.20(4.98,8.70)	5.66(4.31,7.96)	4.27(2.77,6.24)	< 0.0001
METS-IR	51.48(46.41,58.72)	46.34(43.55,49.68)	52.70(48.12,57.77)	59.68(51.50,68.79)	< 0.0001
HOMA-IR	4.08(2.64,6.25)	3.06(2.13,4.39)	4.15(2.77,5.89)	5.41(3.58,8.65)	< 0.0001
Systemic Inflammatory markers
NLR	2.06(1.56,2.67)	1.61(1.28,2.00)	2.04(1.64,2.53)	2.67(2.15,3.44)	< 0.0001
PLR	2.09(1.99,2.18)	2.06(1.98,2.16)	2.09(2.00,2.18)	2.10(2.00,2.21)	< 0.0001
MLR	0.27(0.21,0.33)	0.23(0.19,0.28)	0.27(0.22,0.33)	0.32(0.25,0.42)	< 0.0001
SII	2.70(2.55,2.84)	2.57(2.46,2.69)	2.70(2.57,2.81)	2.83(2.69,2.95)	< 0.0001
SIRI	1.08(0.73,1.57)	0.73(0.58,0.98)	1.09(0.80,1.39)	1.71(1.22,2.30)	< 0.0001

Abbreviations: ISP insulin resistance-systemic inflammation-psychological burden, BMI body mass index, PIR poverty income ratio, DM diabetes mellitus, FBG fasting blood glucose, HbA1c hemoglobin A1c, TG triglyceride, TC total cholesterol, LDL-C low-density lipoprotein cholesterol, HDL-C high-density lipoprotein cholesterol, FLI Fatty liver index, TyG triglyceride-glucose index, CHG cholesterol, high-density lipoprotein, and glucose, LAP lipid accumulation product, eGDR estimated glucose disposal rate, METS-IR metabolic score for insulin resistance, HOMA-IR homeostasis model assessment of insulin resistance, NLR neutrophil-to-lymphocyte ratio, PLR platelet-to-lymphocyte ratio, MLR monocyte-to-lymphocyte ratio, SII systemic immune-inflammation index, SIRI systemic inflammation response index

Higher ISP levels characterized a more adverse multidimensional risk phenotype. Compared with participants in the lower ISP tertiles, those in the higher ISP tertiles were more likely to be female, non-Hispanic White, and current smokers and to have lower educational attainment and a lower poverty-income ratio. They were also more likely to have hypertension, depressive symptoms, and cancer. Biochemically, higher ISP tertiles were accompanied by higher FBG, HbA1c, FLI, insulin resistance surrogates, and systemic inflammation indices, whereas eGDR was lower. These patterns indicate that the ISP index reflects not only its three-component domains but also a broader cardiometabolic and inflammatory risk profile.

Association between the ISP index and mortality risk

Weighted Kaplan‒Meier curves revealed significant separation of survival probabilities across ISP tertiles for both ACM and CVM (all P < 0.001), with progressively lower survival among participants with higher ISP levels (Fig. S3). After participants who died within the first year of follow-up were excluded, the Kaplan‒Meier patterns remained consistent, suggesting that the observed associations were not driven solely by early death or reverse causality (Fig. S4).

In weighted multivariable Cox models, a higher ISP index was consistently associated with increased mortality risk (Table 2). In the fully adjusted model, participants in the highest ISP tertile had a 1.63-fold higher risk of ACM (HR = 1.63, 95% CI: 1.12–2.38; P for trend = 0.01) and a 2.88-fold higher risk of CVM (HR = 2.88, 95% CI: 1.36–6.10; P for trend = 0.003) than participants in the lowest tertile did. When the ISP index was modeled continuously, each 1-SD increase was associated with a 68% higher risk of ACM and an 80% higher risk of CVM. Furthermore, we assessed the associations of the METS-IR, SIRI, and PHQ-9 scores with mortality (Table S3). In the fully adjusted models, when analyzed continuously per 1-SD increase, the SIRI was significantly associated with both ACM and CVM, whereas the METS-IR and the PHQ-9 score were not significantly associated with either outcome. In the tertile-based analyses, METS-IR was not significantly associated with either outcome, whereas the associations of the SIRI and the PHQ-9 score varied according to the outcome and the specific comparison considered. In contrast, the ISP index showed relatively stable positive associations with both ACM and CVM across the continuous and tertile-based analyses.

Table 2.

Associations of the ISP Index with mortality risk among participants with MASLD

ISP index	Model 1		Model 2		Model 3
	HR(95%CI)	P-value	HR(95%CI)	P-value	HR(95%CI)	P-value
All-cause mortality
Per 1-SD increase	1.70(1.50–1.93)	< 0.0001	1.70(1.48–1.94)	< 0.0001	1.68(1.45–1.95)	< 0.0001
Tertile 1	Ref		Ref		Ref
Tertile 2	1.38(1.00–1.91)	0.049	1.46(1.05–2.05)	0.026	1.31(0.90–1.90)	0.155
Tertile 3	2.07(1.50–2.86)	< 0.0001	1.96(1.40–2.73)	< 0.0001	1.63(1.12–2.38)	0.011
P for trend	< 0.0001		< 0.001		0.01
Cardiovascular mortality
Per 1-SD increase	1.89(1.59–2.23)	< 0.0001	1.86(1.56–2.22)	< 0.0001	1.80(1.48–2.20)	< 0.0001
Tertile 1	Ref		Ref		Ref
Tertile 2	1.97(0.89–4.37)	0.09	2.10(0.94–4.68)	0.07	2.02(0.88–4.62)	0.096
Tertile 3	3.60(1.90–6.84)	< 0.0001	3.31(1.70–6.44)	< 0.001	2.88(1.36–6.10)	0.006
P for trend	< 0.0001		< 0.001		0.003

Model 1: crude model;

Model 2: adjusted for sex and age;

Model 3: adjusted for sex, age, race, PIR, marital status, educational levels, BMI, smoking status, alcohol consumption, hypertension, LDL-C, DM, and cancers

Abbreviations: MASLD metabolic dysfunction-associated steatotic liver disease, ISP insulin resistance-systemic inflammation-psychological burden, HR hazard ratio, CI confidence interval, Ref reference, PIR poverty income ratio, BMI body mass index, LDL-C low-density lipoprotein cholesterol, DM diabetes mellitus

Subgroup analyses stratified by sex, age, and diabetes status did not reveal significant multiplicative interactions (all P values for interactions > 0.05), supporting the broad consistency of the association between the ISP index and mortality across major clinical strata (Fig. S5).

Sensitivity analyses of the association between the ISP index and mortality risk

We performed several sensitivity analyses to examine the robustness of the associations between the ISP index and mortality risk. After participants who died within the first year of follow-up were excluded, the positive associations of the ISP index with both ACM and CVM remained generally consistent (Table S4). Similar findings were observed in the PSM population (Table S5), in which baseline covariates were well balanced after matching, as indicated by the reduced standardized mean differences (Fig. S6). In addition, unweighted multivariable Cox regression models yielded results consistent with those of the main survey-weighted analyses (Table S6). Finally, we constructed an equal-weighted ISP index without applying coefficient-derived differential weights and further evaluated its associations with ACM and CVM. The directions and overall patterns of the associations between the equal-weighted ISP index and mortality outcomes were generally consistent with those observed for the original coefficient-weighted ISP index (Table S7). Across these sensitivity analyses, higher ISP levels, particularly the highest ISP tertile, were consistently associated with increased risks of all-cause and cardiovascular mortality. These findings support the robustness of the observed associations, although the results should still be interpreted cautiously given the relatively limited number of cardiovascular mortality events.

Dose‒response relationships between the ISP index and mortality

RCS analyses were conducted to evaluate the dose‒response relationships between the ISP index and all-cause and cardiovascular mortality. The results demonstrated linear dose‒response associations between the ISP index and both all-cause and cardiovascular mortality in the overall population (all P for nonlinearity > 0.05), indicating that the risks of all-cause and cardiovascular mortality gradually increased with increasing ISP. In addition, the dose‒response relationship between the ISP index and cardiovascular mortality remained linear across subgroups stratified by sex, age, and DM status. However, the association between the ISP index and ACM was nonlinear among male participants and older adults (Fig. 2).

Fig. 2.

RCS analysis of the association between the ISP index and mortality risk. A Dose‒response relationship between the ISP index and all-cause mortality in the overall population and stratified by sex, age, and DM status; B dose‒response relationship between the ISP index and cardiovascular mortality in the overall population and stratified by sex, age, and DM status. The solid lines represent estimated hazard ratios, and the shaded areas indicate 95% confidence intervals. Abbreviations: ISP, insulin resistance, systemic inflammation, and psychological burden index; DM, diabetes mellitus; HR, hazard ratio

Incremental predictive value of the ISP index

Using 5-fold cross-validation, we compared Cox models with and without the ISP index for 10-year ACM and CVM prediction. Risk stratification based on out-of-fold predicted risks revealed that participants in the highest risk tertile had significantly higher ACM and CVM risks than those in the lowest tertile in both models (P < 0.001; Fig. 3A). Sankey diagrams show changes in the model-based risk tertiles after the ISP index was added, providing a descriptive visualization of changes in participants’ relative risk classification (Fig. 3B).

Fig. 3.

Incremental predictive value of the ISP index for predicting ACM and CVM. A Kaplan‒Meier survival curves according to tertiles of risk scores derived from Cox models with and without the ISP index for ACM and CVM. B Risk reclassification plots comparing Cox models with and without the ISP index for ACM and CVM. Participants were classified into low-, intermediate-, and high-risk groups according to tertiles of model-derived risk scores. These tertiles represent research-oriented risk strata rather than clinically validated treatment thresholds. C Continuous net reclassification improvement (NRI) with 95% confidence intervals after the ISP index was added to the base model. D Integrated discrimination improvement (IDI) after the ISP index was added to the base model. Abbreviations: ACM, all-cause mortality; CVM, cardiovascular mortality; ISP, insulin resistance, systemic inflammation, and psychological burden index; NRI, net reclassification improvement; IDI, integrated discrimination improvement

The continuous NRIs (Fig. 3C) and IDIs (Fig. 3D) were subsequently calculated using the continuous out-of-fold 10-year predicted risks. With respect to ACM, adding the ISP index yielded a continuous NRI of 0.114 (95% CI: 0.025–0.203) and an IDI of 0.005. For the CVM, the corresponding continuous NRI and IDI were 0.224 (95% CI: 0.035–0.413) and 0.003, respectively (all P < 0.05). Although these improvements were statistically significant, the small absolute IDI values indicated only limited improvement in average discrimination. Therefore, these findings support a modest and exploratory incremental predictive value of the ISP index rather than a substantial improvement in overall model performance or established clinical utility.

ISP index-based machine learning prediction and interpretability

On the basis of 5-fold cross-validation, we evaluated the predictive performance of eight ML models using out-of-fold predicted risk. Compared with the conventional Cox model, the GBM model yielded slightly higher C-indices for both the ACM and the CVM (ACM: 0.844 vs. 0.837, P = 0.004; CVM: 0.854 vs. 0.841, P = 0.015; Table 3). We subsequently assessed the 10-year predictive performance of all the models. Among the evaluated models, GBM showed favorable overall internal performance and was selected for further interpretation, although its differences from close alternatives were limited. Specifically, for ACM prediction, the GBM model achieved a 10-year AUC of 0.869 (95% CI: 0.845–0.892), a Brier score of 0.086, and the highest estimated net benefit across the 20%–80% threshold range in the DCA (Fig. 4A–C). For CVM prediction, the GBM model achieved a 10-year AUC of 0.883 (95% CI: 0.846–0.920), a Brier score of 0.034, and the highest estimated net benefit across the evaluated threshold range in the DCA (Fig. 5A–C). Although the differences in C-indices between the GBM and Cox models were statistically significant, the absolute gains were small, suggesting limited clinical relevance. In addition, given the limited number of CVM events, the AUC estimate for the CVM may be unstable and should be interpreted cautiously. All reported model performance evaluation results were derived from internal validation and should not be interpreted as evidence that the model is ready for clinical deployment.

Table 3.

Comparison of C-indices between machine learning models and the Cox Model

Variables	C-index of all-cause mortality	P-values	C-index of cardiovascular mortality	P-values
Cox	0.837(0.816–0.860)	Ref	0.841(0.798–0.884)	Ref
RSF	0.841(0.821–0.861)	0.482	0.852(0.817–0.887)	0.294
GBM	0.844(0.822–0.866)	0.004	0.854(0.811–0.897)	0.015
CoxBoost	0.829(0.807–0.851)	0.002	0.835(0.792–0.878)	0.133
SurvivalSVM	0.805(0.781–0.829)	< 0.001	0.774(0.729–0.819)	< 0.001
XGBoost	0.844(0.822–0.866)	0.222	0.853(0.812–0.894)	0.224
SuperPC	0.758(0.731–0.785)	< 0.001	0.771(0.718–0.824)	< 0.001
PLSR-Cox	0.835(0.813–0.857)	0.008	0.839(0.794–0.884)	0.448

The optimal hyperparameters for each machine learning model are listed below

Hyperparameters for all-cause mortality:

RSF: ntree = 400, nodesize = 50, mtry = 4, nsplit = 3, splitrule = "logrankscore"

GBM: n.trees = 600, interaction.depth = 3, n.minobsinnode = 10, shrinkage = 0.005

CoxBoost: stepno = 8, penalty = 40

SurvivalSVM: gamma.mu = 0.25, opt.meth = "quadprog", diff.meth = "makediff3", kernel = "lin_kernel", sgf.sv = 3, sigf = 6, maxiter = 30, margin = 0.1, bound = 30

XGBoost: objective = "survival:cox", booster = "gbtree", eval_metric = "cox-nloglik", eta = 0.01, max_depth = 3, subsample = 0.6, colsample_bytree = 0.4, gamma = 0.0, min_child_weight = 5, lambda = 5, alpha = 0

SuperPC: n.components = 4, min.features = 4. PLSR-Cox: nt = 3

Hyperparameters for cardiovascular mortality:

RSF: ntree = 50, nodesize = 100, mtry = 10, nsplit = 1, splitrule = "logrankscore"

GBM: n.trees = 650, interaction.depth = 5, n.minobsinnode = 50, shrinkage = 0.001

CoxBoost: stepno = 7, penalty = 50

SurvivalSVM: gamma.mu = 0.30, opt.meth = "quadprog", diff.meth = "makediff3", kernel = "lin_kernel", sgf.sv = 3, sigf = 5, maxiter = 35, margin = 0.1, bound = 30

XGBoost: objective = "survival:cox", booster = "gbtree", eval_metric = "cox-nloglik", eta = 0.01, max_depth = 3, subsample = 0.6, colsample_bytree = 0.4, gamma = 0.0, min_child_weight = 5, lambda = 5, alpha = 0

SuperPC: n.components = 4, min.features = 4. PLSR-Cox: nt = 3

Abbreviations: RSF random survival forest, GBM gradient boosting machine, SurvivalSVM survival support vector machine, SuperPC supervised principal components, PLSR-Cox partial least squares regression for Cox models

Fig. 4.

Development, evaluation, and interpretation of a machine learning model for predicting ACM. A Time-dependent ROC curves and AUCs of the candidate models for 10-year ACM prediction. B Calibration curves of the candidate models for 10-year ACM prediction. C Decision curve analysis of the candidate models for 10-year ACM prediction. D SHAP beeswarm plot showing the contribution of predictors in the GBM model for ACM prediction. E Individual SHAP explanation plot from the GBM model, showing how each predictor contributed to the predicted risk for a representative participant. F SHAP dependence plot showing the relationship between ISP and SHAP values in the GBM model. Abbreviations: ACM, all-cause mortality; AUC, area under the curve; DCA, decision curve analysis; SHAP, shapley additive explanations; GBM, gradient boosting machine; RSF, random survival forest; ISP, insulin resistance, systemic inflammation, and psychological burden index; DM, diabetes mellitus; PIR, poverty income ratio; LDL-C, low-density lipoprotein cholesterol; BMI, body mass index

Fig. 5.

Development, evaluation, and interpretation of a machine learning model for predicting the CVM. A Time-dependent ROC curves and AUCs of the candidate models for 10-year CVM prediction. B Calibration curves of the candidate models for 10-year CVM prediction. C Decision curve analysis of the candidate models for 10-year CVM prediction. D SHAP beeswarm plot showing the contribution of predictors in the GBM model for CVM prediction. E Individual SHAP explanation plot from the GBM model, showing how each predictor contributed to the predicted risk for a representative participant. F SHAP dependence plot showing the relationship between ISP and SHAP values in the GBM model. Abbreviations: CVM, cardiovascular mortality; AUC, area under the curve; DCA, decision curve analysis; SHAP, Shapley additive explanations; GBM, gradient boosting machine; RSF, random survival forest; ISP, insulin resistance, systemic inflammation, and psychological burden; DM, diabetes mellitus; PIR, poverty income ratio; LDL-C, low-density lipoprotein cholesterol; BMI, body mass index

To interpret the GBM predictions, SHAP summary plots, individual explanation plots, and ISP index dependence plots were generated for the ACM and CVM (ACM: Fig. 4D–F; CVM: Fig. 5D–F). Age showed the greatest feature importance in both models, followed by the ISP index. The dependence plots indicated that higher ISP index values were generally associated with higher SHAP values and therefore contributed to higher model-predicted mortality risks. The individual explanation plots further illustrate how the direction and magnitude of contributions from the ISP index and other variables varied across individual predictions.

Finally, we developed an exploratory web-based calculator based on the GBM model to visualize model-estimated mortality risk at the individual level among adults with MASLD (https://zhongerxiongdiwudi.shinyapps.io/ISPindex/). This calculator is intended solely as a research prototype for risk visualization and hypothesis generation (Fig. S7). It has not been externally validated or recalibrated and should not be regarded as a clinically validated decision-support tool.

Discussion

In this longitudinal cohort study based on NHANES data, we constructed the ISP index by integrating the METS-IR, the SIRI, and the PHQ-9 score to reflect the concurrent metabolic, inflammatory, and depressive symptom burden among individuals with MASLD. A higher ISP index was associated with increased risks of ACM and CVM, and dose–response analyses further revealed that mortality risk generally increased with increasing ISP index. In internal 5-fold cross-validation, the GBM model incorporating the ISP index and other readily available variables showed reasonable discrimination for 10-year ACM and CVM.

MASLD is increasingly understood as a systemic metabolic disorder with cardiovascular complications, and cardiovascular disease is a leading cause of death in patients with fatty liver disease [4, 5]. The present findings extend this concept by suggesting that the joint burden of insulin resistance, systemic inflammation, and depressive symptoms may characterize a high-risk cardiometabolic phenotype within MASLD. This distinction is important because cardiovascular disease is not merely an extrahepatic complication but also a central determinant of prognosis in MASLD. In this context, the ISP index appears to capture systemic cardiometabolic vulnerability rather than hepatic steatosis severity alone.

Comparisons with individual components help clarify the novelty and clinical meaning of the ISP index. METS-IR, the SIRI, and the PHQ-9 score are not new markers, and each has biological relevance. However, they reflect different and only partially overlapping domains of MASLD-related systemic risk. METS-IR mainly captures insulin resistance and lipid-metabolic burden, the SIRI reflects systemic immune-inflammatory activation, and the PHQ-9 score captures depressive symptom burden, which may be linked to neuroendocrine activation, inflammatory signaling, lifestyle behaviors, sleep disturbance, and treatment adherence. Evaluating these markers separately may therefore miss the clinically important situation in which metabolic dysfunction, inflammatory activation, and psychological burden coexist. Accordingly, the value of the ISP index should be understood as integrative risk phenotyping rather than simple statistical superiority over its components. In other words, the ISP index is not intended to replace the METS-IR, SIRI, or PHQ-9 score; instead, it may provide a exploratory framework for characterizing individuals with MASLD in whom insulin resistance, systemic inflammation, and depressive symptom burden jointly define a higher-risk cardiometabolic phenotype.

This integrated interpretation is aligned with recent conceptual shifts in MASLD and cardiovascular–kidney–metabolic health. The MASLD nomenclature emphasizes metabolic dysfunction as the defining context of steatotic liver disease, and contemporary clinical guidance highlights the need to assess cardiometabolic risk in MASLD rather than focusing only on hepatic steatosis [1, 18]. Similarly, the cardiovascular–kidney–metabolic framework emphasizes that metabolic, renal, and cardiovascular abnormalities frequently develop as interconnected processes rather than as isolated conditions [19]. From this perspective, the ISP index may provide a research-oriented summary of multisystem vulnerability.

In the context of changing lifestyles, the growing global burden of obesity and diabetes, and population aging, metabolic dysfunction has become increasingly relevant to adverse cardiovascular and cerebrovascular outcomes [19, 40, 41]. Recent studies have suggested that incorporating metabolic biomarkers into established risk models or integrating complementary risk domains into composite indices may improve risk stratification for these adverse outcomes [42–44]. Similarly, in the present study, we integrated the METS-IR, the SIRI, and the PHQ-9 score to construct the ISP index, which is associated with mortality among individuals with MASLD. Because the number of outcome events was limited, particularly for cardiovascular mortality, we did not further divide the sample into separate derivation and validation subsets. However, marker selection, coefficient-derived weight estimation, association analyses, prediction model development, and performance evaluation were all conducted within the same NHANES analytical sample, which may have introduced optimism bias. Therefore, the observed association strengths and model performance may have been overestimated. An additional concern is the partial component overlap between the definition of the study population and the ISP index. METS-IR incorporates BMI, fasting blood glucose, TG, and HDL-C. BMI and TG are also components of the FLI, while adiposity and glycemic abnormalities form part of the cardiometabolic criteria used to define MASLD. Therefore, METS-IR is not entirely independent of the definition used to identify the MASLD population in this study. Identifying hepatic steatosis using the FLI and applying cardiometabolic criteria to define MASLD may introduce partial mathematical coupling between the ISP index and MASLD classification. Such coupling may enrich the analytical sample for metabolic abnormalities already captured by METS-IR and thereby potentially influence or overestimate the observed association between the ISP index and mortality. Importantly, this potential bias arises primarily from the definition and selection of the study population rather than from direct mathematical overlap between the ISP index and the mortality outcome, because mortality was ascertained independently through National Death Index linkage. Accordingly, the findings should be interpreted specifically within the context of FLI-defined MASLD and require further validation in cohorts in which hepatic steatosis is established using gold-standard methods. Notably, the transportability of the fixed ISP weights beyond the NHANES also remains uncertain. Xie and colleagues reported that the UK Biobank and ESTHER cohorts used different proteomic measurement platforms, resulting in relative expression values that were not directly comparable. Consequently, the coefficients derived from the UK Biobank could not be directly applied to the ESTHER cohort, necessitating cohort-specific coefficient re-estimation; nevertheless, calibration remained suboptimal at higher predicted risks in some groups [45]. These findings illustrate that data-derived coefficients and model performance may vary according to measurement methods, study populations, and application settings. Accordingly, the current ISP weights should not be assumed to be universally transportable and may require cohort-specific re-estimation, whereas prediction models incorporating the ISP index may require recalibration before application to other populations. Future studies should derive the ISP weights in one cohort and validate them in larger, fully independent prospective cohorts, an approach adopted in comparable composite-marker research to reduce optimism in the estimates and assess model transportability [46, 47]. Moreover, the small IDI values and modest absolute improvement of GBM over Cox regression indicate that the current prediction findings remain exploratory. Therefore, the online calculator should be regarded as an exploratory tool for risk visualization rather than a tool that can be directly used for clinical risk screening or decision support.

Notably, medication use may have influenced the observed associations. Antidepressant treatment may lower PHQ-9 scores among treated individuals, but it may also indicate more severe or recurrent depressive symptoms, thereby introducing confounding by indication. Lipid-lowering therapy may reduce TG, LDL-C, and cardiovascular risk, potentially attenuating or modifying associations involving METS-IR and the ISP index. Antidiabetic medications may influence glycemic status, insulin resistance-related phenotypes, and mortality risk. These considerations are clinically important because the ISP index incorporates metabolic, inflammatory, and psychological domains that may all be affected by treatment. In addition, anti-inflammatory, glucocorticoid, or immunomodulatory treatments may alter leukocyte-based inflammatory indices such as the SIRI. However, detailed information on medication indication, duration, dose, adherence, and changes during follow-up could not be reliably reconstructed in the present study. Therefore, residual confounding by treatment status cannot be excluded, and future studies with more detailed longitudinal medication data are needed to further clarify these potential influences.

Study strengths and limitations

This study has several strengths. First, we proposed a novel composite index integrating insulin resistance, systemic inflammation, and psychological burden, providing a new perspective for characterizing multidimensional cardiometabolic risk phenotypes among individuals with MASLD. Second, survey-weighted Cox regression, restricted cubic splines, subgroup analyses, and multiple sensitivity analyses were used to evaluate the associations, allowing the consistency and robustness of the findings to be examined from different perspectives. Finally, multiple machine-learning algorithms were compared using internal 5-fold cross-validation, and SHAP analysis improved the interpretability of the selected model. Although the interactive online calculator remains an initial research prototype, it may provide a foundation for its further development, external validation, and evaluation.

Several limitations should also be acknowledged. First, the lack of independent external validation remains a central limitation. The ISP formula and prediction models were derived and internally evaluated within the same data source. Marker selection using MDA and coefficient-derived weighting were outcome-informed, and the absence of independent external validation, split-sample derivation and validation, or nested validation specifically for ISP index derivation means that the magnitude of the observed associations and incremental prediction results may be optimistic. Moreover, the fixed ISP weights are data derived and may not remain unchanged across populations; the weights may require re-estimation, and prediction models incorporating ISP may require recalibration before application in other populations. Second, in this study, the FLI was used to define hepatic steatosis and, together with cardiometabolic risk criteria, to identify individuals with MASLD. Some variables used in the FLI and MASLD cardiometabolic risk criteria overlap with the inputs of METS-IR, the metabolic component of the ISP index, including TG, BMI, HDL-C, and glucose. This shared input structure may create circularity and mechanically inflate the observed associations between ISP and mortality; therefore, these associations should be interpreted with caution. Third, the tertile-based risk categories used for model comparison and visualization were derived internally from model-predicted risks and should not be interpreted as clinically established thresholds. The numerically small IDI values and modest absolute improvement of GBM over Cox regression further indicate that the current prediction findings remain exploratory and do not support clinical implementation. Fourth, cardiovascular mortality events were relatively limited, which may have reduced precision, widened confidence intervals, and increased the risk of model instability. Fifth, the observational design precludes causal inference. Although we adjusted for multiple demographic, socioeconomic, lifestyle, clinical, and laboratory covariates, residual confounding cannot be excluded. Sixth, although we compared the Cox associations of the ISP index with those of the METS-IR, the SIRI, and the PHQ-9 score, formal incremental prediction models separately adding each individual component to the base model were not developed in the current analysis. Seventh, the PHQ-9 captures depressive symptoms rather than comprehensive psychological burden and should not be interpreted as a psychiatric diagnosis. Finally, the online calculator is a research prototype and should not be used for clinical decision-making without external validation, recalibration, or clinical impact assessment. These limitations point to several directions for future research. External cohorts with imaging-, elastography-, or histology-defined MASLD should be used to validate the ISP index and recalibrate the prediction models.

Conclusions

The ISP index integrates insulin resistance, systemic inflammation, and psychological burden into a single biologically interpretable cardiometabolic risk phenotype index for adults with MASLD. Higher ISP index values were associated with increased risks of both ACM and CVM. The GBM model incorporating the ISP index and other readily available variables provided individualized model-based estimates of mortality risk, and an exploratory online calculator was subsequently developed as a research prototype to visualize these estimates. However, developing and evaluating the ISP index and prediction models within the same NHANES sample, together with the overlap between the input variables of METS-IR and the FLI-based identification and cardiometabolic criteria of MASLD, may have resulted in optimistic estimates of the observed associations and model performance. Overall, the ISP index may support risk phenotyping and preliminary risk enrichment in MASLD.

Future studies should externally validate the ISP index in larger, independent prospective cohorts using assessments of hepatic steatosis and fibrosis that are independent of FLI and compare the current formula with prespecified equal-weight and component-specific alternatives. Longitudinal studies incorporating more comprehensive clinical and biomarker data are needed to evaluate changes in the ISP index and further clarify the observed associations. Recalibration and clinical impact studies are required before the ISP index or online calculator can be considered for clinical practice.

Authors’ contributions

Yan Chen: Conceptualization, Writing—original draft, Writing—review & editing, Methodology, Investigation, Visualization, Data curation. Kailing Xie: Writing—review & editing. Ningkun Xiao: Writing—review & editing, Data curation. Haonan Ju: Visualization, Data curation, Methodology. Qingzhi Zhou: Data curation and Investigation. Xin Zhao: Writing—review & editing, Funding acquisition.

Funding

This study was funded by the National Natural Science Foundation of China (No.82370438).

Data availability

The data used in this study are publicly available from the following website: https://wwwn.cdc.gov/nchs/nhanes/default.aspx

Declarations

Ethics approval and consent to participate

The studies involving human participants received approval from the Research Ethics Review Board of the NCHS. Informed consent was obtained from all the participants. This study complied with the ethical requirements of the Declaration of Helsinki and its subsequent revisions.

Competing interests

The authors declare no competing interests.