A Shared EGR1–IL-32–Macrophage Polarisation Signature in Rheumatoid Arthritis and Metabolic Dysfunction-Associated Steatotic Liver Disease Authors: Lijun Pang, Yunfei Li, Lili Cheng, Chuanbing Huang Published: 2026 ABSTRACT AbstractBackgroundRheumatoid arthritis (RA) and metabolic dysfunction-associated steatotic liver disease (MASLD) frequently co-occur, but their shared molecular and immunometabolic features remain incompletely defined. We integrated population-level epidemiology, hospital-based clinical phenotyping, cross-disease transcriptomic analyses, and patient-level experimental data to systematically characterise RA–MASLD comorbidity.MethodsThe associations of RA with MASLD and hepatic fibrosis-related phenotypes were examined in 5615 adults from NHANES 2017–March 2020 using survey-weighted regression. A retrospective hospital cohort of 100 patients with RA, including 60 with RA only and 40 with RA+MASLD, was used for clinical phenotyping. Nine GEO datasets were analysed independently. Shared genes were identified by sequentially intersecting two hepatic MASLD datasets with a primary RA synovial cohort, followed by direction-concordance filtering, machine-learning prioritisation, external expression assessment, immune-signature analysis, and single-cell localisation. Serum IL-32 and peripheral blood mononuclear cell EGR1 mRNA were measured in 45 participants. Western blotting and flow cytometry were performed as exploratory analyses in a randomly selected subset of nine participants, with three participants per group.ResultsAfter full adjustment, RA was not independently associated with MASLD or the assessed hepatic phenotypes in NHANES (MASLD odds ratio, 1.03; 95% confidence interval, 0.57–1.87). In the hospital cohort, patients with RA+MASLD had higher disease activity, greater hepatocellular injury, and a higher noninvasive fibrosis burden than those with RA alone. Cross-disease transcriptomic analysis identified six concordantly dysregulated genes—IL32, DHRS9, MOXD1, EGR1, RRM2, and PRC1—with EGR1 prioritised for subsequent analyses. EGR1 showed an apparent area under the curve of 0.982 in GSE135251 and an area under the curve of 0.989 in the external hepatic cohort GSE89632. EGR1 expression was inversely associated with the M2 macrophage signature score and serum IL-32 concentration. From healthy controls to the RA-only and RA+MASLD groups, EGR1 mRNA showed a stepwise decrease and serum IL-32 progressively increased. The exploratory flow-cytometric subset showed a corresponding M1-skewed macrophage polarisation pattern.ConclusionAlthough the population-level associations between RA and hepatic phenotypes were attenuated after cardiometabolic adjustment, the hospital cohort, transcriptomic analyses, and patient-level experimental data collectively supported a coordinated EGR1–IL-32–macrophage polarisation signature associated with RA–MASLD comorbidity. These observational findings require prospective replication and functional validation.Keywords: rheumatoid arthritis, metabolic dysfunction-associated steatotic liver disease, EGR1, IL-32, macrophage polarisation, transcriptomic integration Introduction Rheumatoid arthritis (RA) is a chronic systemic autoimmune disease characterised by persistent synovitis, progressive joint destruction, and extra-articular manifestations affecting multiple organ systems.1 The 2021 Global Burden of Disease study estimated that 17.6 million people worldwide were living with RA in 2020 — a 14.1% rise in prevalence since 1990 — with further increases projected through 2050.2 Despite major advances in disease-modifying antirheumatic drug (DMARD) and biologic therapy, RA carries a substantially elevated risk of cardiovascular, metabolic, and hepatic comorbidities not fully explained by traditional risk factors or treatment-related toxicity.3,4 Metabolic dysfunction-associated steatotic liver disease (MASLD) — formally adopted by an international multi-society Delphi consensus in 2023 to replace non-alcoholic fatty liver disease (NAFLD) — is now the leading chronic liver disease worldwide, affecting approximately 30% of adults globally.5,6 Defined by hepatic steatosis in the presence of at least one cardiometabolic risk factor, MASLD encompasses a spectrum from simple steatosis to metabolic dysfunction-associated steatohepatitis (MASH), advanced fibrosis, cirrhosis, and hepatocellular carcinoma.7 Beyond its hepatic consequences, MASLD is a systemic immunometabolic disorder strongly linked to insulin resistance, type 2 diabetes, and cardiovascular disease.8 In patients with RA, concomitant MASLD may complicate hepatic risk assessment and treatment selection, particularly regarding methotrexate use, while MASLD itself is associated with increased liver-related and cardiovascular risk, highlighting the importance of improved risk stratification.7,9,10 RA and MASLD frequently co-occur, although the extent to which this overlap reflects shared cardiometabolic risk factors, treatment-related effects, or common biological processes remains uncertain. In a NHANES-based analysis, the age-adjusted prevalence of MASLD among US adults with RA was approximately 35% under a conservative steatosis threshold (CAP ≥ 285 dB/m),11 while a European population-based cohort reported a higher prevalence in patients with RA than in controls (41.2% vs 28.5%) and an independent RA–MASLD association after multivariable adjustment.12 Estimates nonetheless vary according to the steatosis definition and study population, and whether the observed association reflects a disease-specific relationship or residual confounding remains unresolved. RA is characterised by systemic inflammation and increased insulin resistance, whereas MASLD involves insulin resistance, chronic inflammation, oxidative stress, and mitochondrial dysfunction.13–15 These overlapping inflammatory and metabolic pathways provide a biological rationale for investigating shared immunometabolic features of RA–MASLD comorbidity. To our knowledge, however, these shared molecular and immunometabolic features have not been systematically characterised. Within this broader context, three lines of evidence provide a biological rationale for examining a shared EGR1–IL-32–macrophage polarisation signature. First, macrophage polarisation — the balance between pro-inflammatory M1 and anti-inflammatory M2 phenotypes — is a central pathogenic node in both diseases: M1-skewed infiltrates correlate with synovial inflammation and disease activity in RA,16,17 while hepatic M1 polarisation drives lipotoxic injury and fibrosis progression in MASLD.18 Second, interleukin-32 (IL-32) is prominently upregulated in RA synovium and synovial fluid, where it correlates with disease activity and induces TNF-α and IL-17,19 and has independently been identified among the most strongly upregulated cytokines in MASH liver and as a candidate circulating NAFLD biomarker.20 Third, early growth response 1 (EGR1), a zinc-finger transcription factor that senses inflammatory and metabolic cues, regulates hepatic lipid metabolism and fibrogenesis,21,22 and functions as a transcriptional gatekeeper of inflammatory enhancers in human macrophages.23 However, whether these independently reported observations converge within a reproducible cross-disease transcriptional signature remains unknown. EGR1 and IL-32 were therefore evaluated as data-driven candidates emerging from integrative analyses rather than as prespecified causal mediators. To address this gap, we conducted a four-tier translational study integrating population-level epidemiology in NHANES (n = 5615), hospital-based phenotyping of a Chinese RA cohort (n = 100), cross-disease analysis of nine GEO datasets with machine-learning-based candidate-gene prioritisation, and patient-level experimental assessment in 45 participants. We aimed to identify shared molecular and immunometabolic features associated with RA–MASLD comorbidity and to evaluate their consistency across transcriptomic, immune, single-cell, and experimental analyses. Materials and Methods Study Design and Overview This study used a four-tier translational design to characterise shared molecular and immunometabolic features of rheumatoid arthritis (RA) and metabolic dysfunction-associated steatotic liver disease (MASLD). Tier 1 examined the population-level association between RA and hepatic phenotypes in a nationally representative US cohort (NHANES 2017–March 2020; n = 5615 adults). Tier 2 characterised the clinical profile of comorbid RA–MASLD in a retrospective Chinese hospital cohort comprising patients with RA only (n = 60) and RA with comorbid MASLD (n = 40). Tier 3 integrated nine GEO transcriptomic datasets, including a single-cell RNA-sequencing dataset of RA synovium, to identify shared dysregulated genes, prioritise candidate genes using machine-learning approaches, and assess their expression in independent cohorts. Tier 4 evaluated key findings in a clinical validation set of 45 participants (15 healthy controls, 15 RA only, and 15 RA+MASLD); ELISA and qRT-PCR were performed in all participants, whereas Western blotting and flow cytometry were conducted in an exploratory subset of nine participants (three per group). The analytical workflow is summarised in Figure 1. Four-tier translational framework. Integrated design used to examine the shared immunometabolic basis of rheumatoid arthritis (RA) and metabolic dysfunction-associated steatotic liver disease (MASLD). Tier 1 tested population-level RA-hepatic associations in NHANES 2017-March 2020 (n = 5615). Tier 2 compared RA only (n = 60) with RA+MASLD (n = 40) in a Chinese hospital cohort. Tier 3 integrated nine GEO transcriptomic datasets to identify six shared genes (IL32, DHRS9, MOXD1, EGR1, RRM2 and PRC1) and prioritise EGR1 for subsequent analyses. Tier 4 assessed the EGR1–IL-32–macrophage polarisation signature in healthy controls, patients with RA only, and patients with RA+MASLD (n = 15 per group), with Western blotting and flow cytometry performed as exploratory analyses in three participants per group. Abbreviations: GEO, Gene Expression Omnibus; MASLD, metabolic dysfunction-associated steatotic liver disease; NHANES, National Health and Nutrition Examination Survey; RA, rheumatoid arthritis. NHANES Discovery Cohort Data were drawn from the NHANES 2017–March 2020 pre-pandemic cycle, a cross-sectional survey of the non-institutionalised US civilian population conducted by the National Center for Health Statistics. All analyses incorporated mobile examination centre weights (WTMECPRP), primary sampling units (SDMVPSU), and stratum identifiers (SDMVSTRA) to generate nationally representative estimates. The protocol was approved by the National Center for Health Statistics Research Ethics Review Board, and all participants provided written informed consent. Of 15,560 enrolled participants, we sequentially excluded those aged under 18 years (n = 5867), without valid transient elastography (n = 1926), with heavy alcohol use (>2 standard drinks/day in women, >3 in men; n = 1308), with unknown RA status (n = 794), or with missing body mass index (BMI; n = 50), leaving 5615 adults for analysis. RA was defined as self-reported physician-diagnosed arthritis (MCQ160A = 1) with the subtype specified as rheumatoid (MCQ195 = 2); participants with other subtypes were classified as non-RA, and those with unclassifiable subtypes were excluded. Hepatic steatosis and stiffness were assessed by vibration-controlled transient elastography (FibroScan®, Echosens, Paris, France), retaining only examinations with a complete acquisition status (LUAXSTAT = 1) and an interquartile range-to-median ratio <30%. MASLD was defined according to the 2023 multi-society Delphi consensus as a controlled attenuation parameter (CAP) ≥248 dB/m together with at least one cardiometabolic risk factor: overweight/obesity (BMI ≥25 kg/m2); physician-diagnosed diabetes or HbA1c ≥5.7%; hypertension (self-reported diagnosis, antihypertensive use, or blood pressure ≥130/80 mmHg); or dyslipidaemia (triglycerides ≥150 mg/dL, or HDL-cholesterol <40 mg/dL in men and <50 mg/dL in women). High fibrosis risk was defined as a liver stiffness measurement (LSM) ≥8.0 kPa, and the FIB-4 index was calculated as (age × AST) / (platelet count × √ALT). Three hierarchical models were prespecified: Model 1 adjusted for age, sex, and race/ethnicity; Model 2 additionally for education, poverty-to-income ratio, and smoking status; and Model 3 additionally for BMI, diabetes, hypertension, and the cardiometabolic composite. Prespecified subgroup analyses stratified the Model 3 estimates by sex, age, obesity, diabetes, and hypertension. Four sensitivity analyses examined stricter outcome thresholds (CAP ≥274 dB/m or LSM ≥9.7 kPa) and the exclusion of underweight (BMI <18.5 kg/m2) or Non-Hispanic Asian participants. Clinical Validation Cohort Consecutive RA patients admitted to the First Affiliated Hospital of Anhui University of Chinese Medicine (Hefei, China) between January and December 2025 were retrospectively screened. RA was diagnosed by the 2010 ACR/EULAR criteria. Eligible patients were aged ≥18 years with complete clinical, laboratory, and hepatic imaging data at the index visit. We excluded significant alcohol consumption (>140 g/week in men, >70 g/week in women), hepatitis B or C infection, other chronic liver disease, active malignancy, pregnancy or lactation, and missing data essential for MASLD or fibrosis assessment. MASLD was diagnosed per the 2023 multi-society Delphi consensus: hepatic steatosis (CAP ≥248 dB/m by vibration-controlled transient elastography or grade ≥1 steatosis on B-mode ultrasonography) plus at least one cardiometabolic risk factor (BMI ≥23 kg/m2 [Asian cut-off], fasting glucose ≥5.6 mmol/L or type 2 diabetes, hypertriglyceridaemia, low HDL-cholesterol, or hypertension), with other steatotic liver aetiologies excluded. Steatosis and stiffness were measured on the FibroTouch® system (Wuxi Hisky Medical Technologies, Wuxi, China), and noninvasive fibrosis was indexed by FIB-4 and APRI [(AST/40) ÷ platelet count × 100]. Fasting venous blood was assayed for liver function, lipid profile, glucose markers, C-reactive protein (CRP), erythrocyte sedimentation rate (ESR), and full blood count. RA disease activity was scored by DAS28-CRP. The study was approved by the Ethics Committee of the First Affiliated Hospital of Anhui University of Chinese Medicine (Approval No. 2023AH-76); informed consent is described in the Declarations. Transcriptomic Datasets Nine publicly available GEO datasets were used for complementary analytic purposes. GSE135251 and GSE126848 were used to derive the MASLD consensus gene set, which was intersected with the RA synovial dataset GSE55457 for primary cross-disease shared-gene identification. GSE77298 served as an independent RA synovial confirmation cohort and was not included in the primary intersection. External validation cohorts comprised GSE89632 (hepatic tissue), GSE48452 (hepatic tissue), and GSE93272 (peripheral whole blood). GSE117769 served as a supportive WGCNA dataset, and GSE200815 provided single-cell resolution of RA synovium. Ensembl gene identifiers were mapped to HGNC symbols using org.Hs.eg.db (v3.21.0). Detailed dataset characteristics and analysis-specific sample subsets are provided in Table S1. Differential Expression and Weighted Gene Co-Expression Network Analysis For RNA-seq datasets, raw count matrices were processed with DESeq2 (v1.48.2)24 after retaining genes with counts ≥10 in at least one-third of the samples included in each analysis; differentially expressed genes (DEGs) were defined as adjusted P < 0.05 and |log2FC| > 1. For the microarray datasets used in differential-expression and shared-gene analyses, expression matrices were retrieved from series matrix files, mapped to gene symbols using the corresponding platform annotations, collapsed to gene level by retaining the probe with the highest across-sample variance, and analysed using limma (v3.64.3)25,26 with empirical Bayes moderation; DEGs were defined as adjusted P < 0.05 and |log2FC| > 0.5. WGCNA was performed independently on GSE135251 and GSE117769 using the WGCNA package (v1.73).27 After variance-stabilising transformation, the 5000 most variable genes were retained as input. Soft-thresholding powers were selected separately for each dataset based on the scale-free topology diagnostics generated by pickSoftThreshold. Signed hybrid co-expression networks were constructed, and modules were detected using the blockwiseModules function (minimum module size = 30; merge cut height = 0.25). Module eigengenes were correlated with clinical traits, including fibrosis stage, NAFLD activity score (NAS), and MASLD status in GSE135251 and RA status in GSE117769. GSE117769 was treated as supportive co-expression evidence and was not used for shared-gene selection. Shared Gene Identification and Functional Annotation A sequential intersection and direction-of-effect filtering strategy was used to identify genes concordantly dysregulated across MASLD and RA. The differentially expressed genes from GSE135251 and GSE126848 were first intersected to derive a 229-gene MASLD consensus set. This set was subsequently intersected with the RA synovial differentially expressed genes from GSE55457, yielding eight candidate shared genes. Six genes (IL32, DHRS9, MOXD1, EGR1, RRM2, and PRC1) were retained after requiring concordant log2 fold-change directions across GSE135251, GSE126848, and GSE55457. GSE77298 was used as an independent RA synovial confirmation cohort and was not included in the primary intersection. Gene Ontology biological process (GO-BP) enrichment of the six shared genes was performed using clusterProfiler (v4.16.0), with org.Hs.eg.db as the annotation database and Benjamini–Hochberg-adjusted P < 0.05 as the significance threshold. Pairwise Pearson correlations among the six shared genes were calculated using variance-stabilised expression values from the full GSE135251 cohort (n = 216). Machine-Learning-Based Candidate-Gene Prioritisation Three machine-learning algorithms were applied to the 229-gene MASLD consensus set in GSE135251 to prioritise a candidate gene. Least absolute shrinkage and selection operator (LASSO) logistic regression was implemented using the glmnet package with 10-fold cross-validation to determine λ.min. Random Forest classification was performed using the randomForest package (ntree = 1000), with variable importance assessed by mean decrease in accuracy. Support vector machine with recursive feature elimination (SVM-RFE) was implemented in caret using a linear kernel and 5-fold cross-validation. The prioritised candidate gene was defined as a gene satisfying two criteria: selection by at least two of the three algorithms and membership in the six directionally concordant RA–MASLD shared genes. Gene set enrichment analysis of the ranked GSE135251 transcriptome was performed against a predefined subset of 19 Hallmark-related gene sets using fgsea (v1.32.0; nPermSimple = 10,000; minSize = 10; maxSize = 500; seed = 42). Gene Ontology biological-process and KEGG pathway enrichment analyses of the 229 consensus genes were conducted using clusterProfiler. External Validation and Exploratory Diagnostic Modelling Three independent GEO cohorts were used for external validation: GSE89632 and GSE48452 for MASLD-related hepatic expression and GSE93272 for RA peripheral-blood expression. Multiple probes mapping to the same gene symbol were collapsed by arithmetic mean, and log2 transformation was applied when the expression range indicated untransformed intensities. Between-group expression differences were assessed using the Wilcoxon rank-sum test. Stage-dependent expression differences across the MASLD histological spectrum in GSE135251 were assessed using the Kruskal–Wallis test, and mild versus advanced fibrosis among NASH samples was compared using the Wilcoxon rank-sum test. Receiver operating characteristic analyses were performed using the pROC package, with area under the curve (AUC) values and 95% confidence intervals calculated for individual genes and logistic-regression models. External discriminatory performance of EGR1 was evaluated in GSE89632 for NASH and in GSE93272 for RA, whereas GSE48452 was used for expression replication. Within GSE135251, apparent discriminatory performance was assessed for each of the six shared genes, a combined six-gene logistic-regression model, and fibrosis-stage classification among NASH samples. An exploratory three-gene logistic-regression model incorporating EGR1, IL32, and RRM2 was visualised as a nomogram using regplot. Internal calibration was assessed using 200 bootstrap resamples with the rms package, and model fit was examined using the Hosmer–Lemeshow test. Because the GSE135251 models were fitted and evaluated in the same highly imbalanced dataset (206 MASLD samples and 10 controls), their AUCs and calibration estimates were interpreted as apparent internal performance rather than independent diagnostic validation. Immune Infiltration and Single-Cell RNA Sequencing Analysis Immune-cell enrichment in GSE135251 was estimated by single-sample gene set enrichment analysis (ssGSEA) implemented in GSVA (v2.2.1), using nine predefined custom marker-gene signatures representing M1 macrophages, M2 macrophages, CD8⁺ T cells, CD4⁺ T cells, natural killer cells, B cells, monocytes, neutrophils, and dendritic cells (Table S3). The resulting scores were interpreted as relative enrichment scores rather than direct estimates of cell proportions. A composite six-gene module score was calculated as the mean variance-stabilised expression of IL32, DHRS9, MOXD1, EGR1, RRM2, and PRC1. Pearson correlation was used to assess associations between the six-gene module score and the M1 macrophage score, between EGR1 expression and the M2 macrophage score, and between individual shared genes and immune-cell scores. Associations between immune-cell scores and ordinal fibrosis stage were assessed using Spearman correlation. Differences across fibrosis stages were evaluated using the Kruskal–Wallis test. Samples were dichotomised at the median EGR1 expression, and immune-cell scores were compared between the EGR1-high and EGR1-low groups using the Wilcoxon rank-sum test. Single-cell RNA-sequencing data from RA synovium (GSE200815) were processed using Seurat (v5.3.1). Cells with ≤200 or ≥6000 detected genes or mitochondrial transcript proportions ≥20% were excluded. After log normalisation, the 2000 most variable genes were identified, and inter-sample variation was addressed using Harmony integration. Graph-based clustering and uniform manifold approximation and projection (UMAP) were performed using the Harmony-corrected dimensions. Cell clusters were manually annotated according to canonical marker expression and consolidated into six major populations: fibroblasts, endothelial cells, T cells, macrophages, B cells, and plasma cells. The distributions of EGR1 and the six shared genes across the annotated cell populations were visualised using violin and dot plots. Experimental Validation in Clinical Samples To corroborate key transcriptomic findings, 45 participants were enrolled: 15 separately recruited healthy controls (HC) and, drawn from the clinical validation cohort, 15 RA-only and 15 RA+MASLD patients (patients randomly sampled within each stratum). Serum IL-32 (ELISA) and EGR1 mRNA in peripheral blood mononuclear cells (PBMCs; qRT-PCR) were measured in all 45 participants; owing to limited material, Western blot and flow cytometric phenotyping were performed as exploratory analyses in a randomly selected subset of nine (three per group). Laboratory assays and data analyses were performed by personnel blinded to group allocation. Serum IL-32 was quantified by sandwich ELISA in duplicate. PBMC EGR1 mRNA was measured by qRT-PCR and calculated by the 2−ΔΔCt method, normalised to β-actin. EGR1 protein was assessed by Western blot with GAPDH as the loading control. Macrophage polarisation was assessed by flow cytometry (CytoFLEX, Beckman Coulter), with M1-like and M2-like macrophages defined as CD45⁺CD68⁺CD86⁺ and CD45⁺CD68⁺CD163⁺, respectively. Full reagent, antibody, and primer details are provided in Supplementary Methods and Table S2. Statistical Analysis All analyses were performed in R (v4.5.1), with two-sided P < 0.05 defining statistical significance. NHANES analyses were survey-weighted using the survey package. Continuous variables are reported as weighted means ± SD or medians (IQR), and categorical variables as weighted percentages with unweighted counts. Between-group comparisons used design-based t-tests and Rao–Scott χ2-tests; survey-weighted logistic and linear regression models estimated associations between RA and hepatic outcomes across three nested models, and subgroup heterogeneity was assessed using Wald tests for multiplicative interaction terms. In the clinical cohort, continuous variables were assessed for normality using the Shapiro–Wilk test and are presented as mean ± SD or median (IQR), as appropriate; categorical variables are presented as frequencies and percentages. Between-group comparisons used independent-samples t-tests, Mann–Whitney U-tests, or Fisher’s exact tests, as appropriate. Factors associated with MASLD were evaluated using Firth-penalised logistic regression with z-standardised continuous predictors, and prespecified clinical associations were assessed using Spearman correlation. For transcriptomic analyses, two-group comparisons used the Wilcoxon rank-sum test and comparisons across multiple disease or fibrosis stages used the Kruskal–Wallis test. WGCNA module–trait associations, the six-gene module score–M1 association, the EGR1–M2 association, and shared-gene–immune-cell associations were assessed using Pearson correlation; associations between immune-cell scores and ordinal fibrosis stage were assessed using Spearman correlation. Diagnostic performance was quantified using AUCs with 95% confidence intervals calculated using the pROC package. Experimental data were compared using one-way ANOVA or Kruskal–Wallis tests, as appropriate, with multiplicity-adjusted post-hoc comparisons. Correlations between EGR1, IL-32, and clinical variables were assessed using Spearman’s rho across all 45 participants or among patients with RA (n = 30), as specified. Results Population-Level Hepatic Differences in RA Were Attenuated After Cardiometabolic Adjustment After prespecified exclusions, 5615 adults from the NHANES 2017–March 2020 pre-pandemic cycle were included, comprising 389 participants with RA and 5226 without RA (Figure 2A; Table 1). Participants with RA were older than those without RA (weighted mean, 59.1 vs 48.7 years; P < 0.001) and had higher weighted prevalences of diabetes (21.8% vs 10.3%) and hypertension (67.3% vs 50.8%; both P < 0.001). They also had a higher median FIB-4 index (1.2 vs 0.9; P < 0.001). Standardised mean-difference analysis demonstrated substantial baseline imbalance, particularly in age, hypertension, diabetes, and FIB-4 (Figure S1). Population-level associations between RA and hepatic phenotypes in NHANES. (A) Participant flow from 15,560 enrolled NHANES participants to the final analytic sample of 5615 adults. (B) Weighted prevalences of MASLD and high fibrosis risk according to RA status. (C) Distributions of CAP and LSM according to RA status; dashed lines indicate the corresponding diagnostic thresholds. (D) Fully adjusted Model 3 associations between RA and hepatic outcomes; odds ratios are shown for binary outcomes and β coefficients for continuous outcomes. (E) Prespecified subgroup analyses of the association between RA and MASLD. (F) Sensitivity analyses using stricter CAP or LSM thresholds and alternative exclusion criteria. Model 3 was adjusted for age, sex, race/ethnicity, education, poverty-to-income ratio, smoking status, BMI, diabetes, hypertension, and cardiometabolic composite. All analyses accounted for NHANES examination weights, strata, and primary sampling units. Boxplots show medians and interquartile ranges; error bars or shaded intervals indicate 95% confidence intervals. Abbreviations: BMI, body mass index; CAP, controlled attenuation parameter; CI, confidence interval; LSM, liver stiffness measurement; MASLD, metabolic dysfunction-associated steatotic liver disease; NHANES, National Health and Nutrition Examination Survey; OR, odds ratio; RA, rheumatoid arthritis. Table 1Weighted Baseline Characteristics of NHANES 2017–March 2020 Pre-Pandemic Participants by RA StatusCharacteristicTotal (N = 5615)ANon-RA (n = 5226)ARA (n = 389)AP valueBAge, years49.2 (17.2)48.7 (17.2)59.1 (13.2)<0.001Sex, n (%)0.691Male2747 (48.6%)2562 (48.5%)185 (50.1%)Female2868 (51.4%)2664 (51.5%)204 (49.9%)Race/ethnicity, n (%)0.025Non-Hispanic White1888 (63.0%)1766 (63.3%)122 (57.8%)Non-Hispanic Black1535 (11.7%)1389 (11.4%)146 (17.4%)Mexican American578 (7.3%)542 (7.3%)36 (6.9%)Other Hispanic544 (7.2%)509 (7.3%)35 (6.4%)Non-Hispanic Asian807 (7.0%)777 (7.1%)30 (4.9%)Other/Multiracial263 (3.8%)243 (3.6%)20 (6.5%)Education level, n (%)0.008 0.05; Figure 2E). Sensitivity analyses using stricter CAP or LSM thresholds and alternative exclusion criteria did not materially change these findings (Figure 2F). Comorbid MASLD Was Associated with Greater Inflammatory and Hepatic Disease Burden in Patients with RA Of 415 patients screened, 100 met the eligibility criteria, including 60 patients with RA only and 40 with RA+MASLD (Figure 3A; Table 2). Patients with RA+MASLD were older and had a higher BMI than those with RA only, whereas sex distribution and RA duration did not differ significantly between the groups. Inflammatory measures were higher in the RA+MASLD group, including DAS28 (4.97 ± 0.61 vs 4.16 ± 0.63), ESR, and CRP (all P < 0.001; Figure 3B). Metabolic and hepatic abnormalities were also more pronounced, with higher fasting plasma glucose, HbA1c, triglycerides, LDL cholesterol, and liver enzyme levels, together with lower HDL cholesterol and albumin levels (all P < 0.001 for these variables; Figure 3C; Table 2). Measures of hepatic steatosis, liver stiffness, and fibrosis risk—including CAP, LSM, FIB-4, and APRI—were also higher in the RA+MASLD group (all P < 0.001; Figure 3D). FIB-4 ≥1.30 was observed in 62.5% of patients with RA+MASLD and 15.0% of those with RA only, while 50.0% and 11.7%, respectively, met the intermediate-to-high APRI threshold (both Fisher’s exact P < 0.001; Figure 3E). Inflammatory, metabolic, and hepatic features of RA with comorbid MASLD. (A) Flow diagram of the Chinese hospital cohort. Among 415 patients with RA screened, 100 met the eligibility criteria and were classified as RA only (n = 60) or RA+MASLD (n = 40). (B) Between-group comparisons of DAS28, ESR, and CRP. (C) Between-group comparisons of hepatic enzymes and metabolic measures. (D) Between-group comparisons of non-invasive steatosis-, liver-stiffness-, and fibrosis-related indices. Comparisons in panels B–D were performed using the independent-samples t-test or Mann–Whitney U-test, as appropriate. (E) FIB-4 and APRI risk categories; Fisher’s exact test. (F) Spearman correlation heatmap of prespecified inflammatory, metabolic, hepatic, and fibrosis-related variables; only statistically significant correlations are displayed. (G) Multivariable Firth-penalised logistic regression analysis of factors associated with MASLD among patients with RA; odds ratios for continuous predictors are expressed per 1-SD increase after z-score standardisation. Boxplots show medians and interquartile ranges with individual observations. Abbreviations: ALB, albumin; ALD, alcohol-related liver disease; ALT, alanine aminotransferase; APRI, aspartate aminotransferase-to-platelet ratio index; AST, aspartate aminotransferase; BMI, body mass index; CAP, controlled attenuation parameter; CI, confidence interval; CLD, chronic liver disease; CRP, C-reactive protein; DAS28, 28-joint Disease Activity Score; ESR, erythrocyte sedimentation rate; FIB-4, fibrosis-4 index; FPG, fasting plasma glucose; GGT, gamma-glutamyl transferase; HbA1c, glycated haemoglobin; HDL-C, high-density lipoprotein cholesterol; LSM, liver stiffness measurement; MASLD, metabolic dysfunction-associated steatotic liver disease; OR, odds ratio; RA, rheumatoid arthritis; SD, standard deviation; TC, total cholesterol; TG, triglycerides. Table 2Baseline Characteristics of Patients with RA with and without Comorbid MASLD in the Chinese Hospital Cohort (n = 100)CharacteristicRA Only (n = 60)RA + MASLD (n = 40)P valueDemographicsAge, years50.68 (9.84)57.00 (9.29)0.002BMI, kg/m223.39 (2.24)30.45 (2.85)<0.001Male sex, n (%)7 (11.7%)8 (20.0%)0.268RA disease characteristicsRA duration, years8.31 (4.51)8.52 (3.63)0.796DAS284.16 (0.63)4.97 (0.61)<0.001ESR, mm/h29.00 [18.75, 34.25]43.00 [32.25, 54.25]<0.001CRP, mg/L5.85 [4.60, 7.90]12.70 [8.92, 16.45]<0.001Metabolic parametersFPG, mmol/L5.25 (0.36)6.11 (0.60)<0.001HbA1c, %5.50 (0.25)5.93 (0.50)<0.001Triglycerides, mmol/L1.11 [0.92, 1.40]1.86 [1.62, 2.37]<0.001Total cholesterol, mmol/L4.92 (0.74)5.28 (0.84)0.034LDL-cholesterol, mmol/L2.70 (0.55)3.22 (0.65)<0.001HDL-cholesterol, mmol/L1.26 (0.19)0.97 (0.16)<0.001Liver function testsALT, U/L46.87 (10.42)61.73 (10.44)<0.001AST, U/L36.65 (9.35)45.85 (8.00)<0.001GGT, U/L33.47 (9.52)49.05 (8.11)<0.001ALP, U/L71.55 (8.50)85.55 (10.50)<0.001TBIL, μmol/L9.80 (2.18)8.94 (1.84)0.036Albumin, g/L43.20 (1.31)41.78 (2.05)<0.001Complete blood countWBC, ×109/L6.48 (1.12)7.54 (1.20)<0.001HGB, g/L128.70 (8.83)120.45 (10.26)<0.001PLT, ×109/L271.37 (54.59)244.30 (45.76)0.009Hepatic imaging and noninvasive fibrosisCAP, dB/m210.83 (13.32)309.15 (17.34)<0.001LSM, kPa5.60 [5.38, 6.00]8.10 [7.50, 8.93]<0.001FIB-41.00 [0.80, 1.17]1.36 [1.17, 1.59]<0.001APRI0.34 [0.26, 0.42]0.50 [0.39, 0.54]<0.001Medications, n (%)Methotrexate41 (68.3%)24 (60.0%)0.402Leflunomide27 (45.0%)21 (52.5%)0.542Hydroxychloroquine26 (43.3%)13 (32.5%)0.303Sulphasalazine9 (15.0%)4 (10.0%)0.555Current glucocorticoid25 (41.7%)24 (60.0%)0.102Targeted therapy (b/tsDMARD)1 (1.7%)8 (20.0%)0.003NSAID12 (20.0%)18 (45.0%)0.013Traditional Chinese medicine27 (45.0%)14 (35.0%)0.407Notes: Continuous variables are presented as mean (SD) for approximately normal distributions or median [interquartile range] for skewed distributions. Categorical variables are presented as n (%). Between-group comparisons used the independent-samples t-test, Mann–Whitney U-test, or Fisher’s exact test, as appropriate. Bold text is used for column and section headings only.Abbreviations: ALP, alkaline phosphatase; ALT, alanine aminotransferase; APRI, AST-to-platelet ratio index; AST, aspartate aminotransferase; b/tsDMARD, biological or targeted synthetic disease-modifying antirheumatic drug; BMI, body mass index; CAP, controlled attenuation parameter; CRP, C-reactive protein; DAS28, 28-joint Disease Activity Score; ESR, erythrocyte sedimentation rate; FIB-4, fibrosis-4 index; FPG, fasting plasma glucose; GGT, gamma-glutamyl transferase; HbA1c, glycated haemoglobin; HDL, high-density lipoprotein; HGB, haemoglobin; LDL, low-density lipoprotein; LSM, liver stiffness measurement; MASLD, metabolic dysfunction-associated steatotic liver disease; NSAID, non-steroidal anti-inflammatory drug; PLT, platelet count; RA, rheumatoid arthritis; TBIL, total bilirubin; WBC, white blood cell count. Baseline Characteristics of Patients with RA with and without Comorbid MASLD in the Chinese Hospital Cohort (n = 100) Notes: Continuous variables are presented as mean (SD) for approximately normal distributions or median [interquartile range] for skewed distributions. Categorical variables are presented as n (%). Between-group comparisons used the independent-samples t-test, Mann–Whitney U-test, or Fisher’s exact test, as appropriate. Bold text is used for column and section headings only. Abbreviations: ALP, alkaline phosphatase; ALT, alanine aminotransferase; APRI, AST-to-platelet ratio index; AST, aspartate aminotransferase; b/tsDMARD, biological or targeted synthetic disease-modifying antirheumatic drug; BMI, body mass index; CAP, controlled attenuation parameter; CRP, C-reactive protein; DAS28, 28-joint Disease Activity Score; ESR, erythrocyte sedimentation rate; FIB-4, fibrosis-4 index; FPG, fasting plasma glucose; GGT, gamma-glutamyl transferase; HbA1c, glycated haemoglobin; HDL, high-density lipoprotein; HGB, haemoglobin; LDL, low-density lipoprotein; LSM, liver stiffness measurement; MASLD, metabolic dysfunction-associated steatotic liver disease; NSAID, non-steroidal anti-inflammatory drug; PLT, platelet count; RA, rheumatoid arthritis; TBIL, total bilirubin; WBC, white blood cell count. Spearman correlation analysis across the 100 patients showed positive interrelationships among inflammatory measures, liver enzymes, and steatosis- and fibrosis-related indices, whereas HDL cholesterol and albumin were inversely correlated with several metabolic and inflammatory measures (Figure 3F). In the multivariable Firth-penalised logistic regression model, BMI, triglycerides, and DAS28 retained associations with MASLD after mutual adjustment. Because continuous predictors were z-score standardised, the adjusted ORs were expressed per 1-SD increase: 35.13 for BMI (95% CI, 4.61–185.49; P < 0.001), 47.10 for triglycerides (95% CI, 3.19–4866.93; P = 0.001), and 6.56 for DAS28 (95% CI, 1.47–12.51; P = 0.010). Age was not associated with MASLD after adjustment (OR, 1.65; 95% CI, 0.45–14.09; P = 0.472) (Figure 3G; Table 3). Among the recorded medication exposures, b/tsDMARD use (20.0% vs 1.7%; P = 0.003) and NSAID use (45.0% vs 20.0%; P = 0.013) were more frequent in the RA+MASLD group, whereas the other recorded therapies did not differ significantly between the groups (Table 2). Table 3Firth-Penalised Logistic Regression Analysis of Factors Associated with MASLD Among Patients with RA in the Chinese Hospital Cohort (n = 100)VariableUnivariable OR (95% CI)P (univariable)Multivariable OR (95% CI)*P (multivariable)Age1.07 (1.02–1.12)0.0021.65 (0.45–14.09)0.472Female sex0.54 (0.18–1.58)0.255——BMI3.00 (2.03–5.35)<0.00135.13 (4.61–185.49)<0.001RA duration1.01 (0.92–1.11)0.802——DAS288.21 (3.61–22.32)<0.0016.56 (1.47–12.51)0.010CRP1.59 (1.35–1.95)<0.001——Triglycerides203.84 (31.65–2436.60)<0.00147.10 (3.19–4866.93)0.001HbA1c27.00 (6.57–154.72)<0.001——Current glucocorticoid use2.07 (0.93–4.69)0.075——Notes: *The multivariable model included age, BMI, DAS28, and triglycerides. Univariable ORs for continuous variables are expressed per original measurement unit, whereas multivariable ORs are expressed per 1-SD increase after z-score standardisation (age, 10.1 years; BMI, 4.3 kg/m2; DAS28, 0.74 units; triglycerides, 0.66 mmol/L). Firth-penalised logistic regression was used to reduce separation and sparse-data bias. Variables marked “—” were not included in the multivariable model. Bold values indicate statistically significant associations in the multivariable model (P < 0.05).Abbreviations: BMI, body mass index; CI, confidence interval; CRP, C-reactive protein; DAS28, 28-joint Disease Activity Score; HbA1c, glycated haemoglobin; MASLD, metabolic dysfunction-associated steatotic liver disease; OR, odds ratio; RA, rheumatoid arthritis; SD, standard deviation. Firth-Penalised Logistic Regression Analysis of Factors Associated with MASLD Among Patients with RA in the Chinese Hospital Cohort (n = 100) Notes: *The multivariable model included age, BMI, DAS28, and triglycerides. Univariable ORs for continuous variables are expressed per original measurement unit, whereas multivariable ORs are expressed per 1-SD increase after z-score standardisation (age, 10.1 years; BMI, 4.3 kg/m2; DAS28, 0.74 units; triglycerides, 0.66 mmol/L). Firth-penalised logistic regression was used to reduce separation and sparse-data bias. Variables marked “—” were not included in the multivariable model. Bold values indicate statistically significant associations in the multivariable model (P < 0.05). Abbreviations: BMI, body mass index; CI, confidence interval; CRP, C-reactive protein; DAS28, 28-joint Disease Activity Score; HbA1c, glycated haemoglobin; MASLD, metabolic dysfunction-associated steatotic liver disease; OR, odds ratio; RA, rheumatoid arthritis; SD, standard deviation. Cross-Disease Transcriptomic Integration Identified Six Genes Concordantly Dysregulated Across MASLD Liver and RA Synovium Differential-expression analysis identified 1299 DEGs in the hepatic MASLD dataset GSE135251, including 793 upregulated and 506 downregulated genes, and 1071 DEGs in GSE126848, including 600 upregulated and 471 downregulated genes (Figure 4A and B). In RA synovium, 417 DEGs were identified in GSE55457 (288 upregulated and 129 downregulated), which served as the primary RA synovial intersection cohort. GSE77298 yielded 462 DEGs (262 upregulated and 200 downregulated) and was analysed separately as an independent RA synovial confirmation cohort (Figure 4C and D). Cross-disease transcriptomic integration and identification of shared genes in MASLD and RA. (A–D) Volcano plots of differentially expressed genes in the hepatic MASLD datasets GSE135251 and GSE126848 and the RA synovial datasets GSE55457 and GSE77298. GSE55457 served as the primary RA synovial intersection cohort, whereas GSE77298 was used as an independent confirmation cohort. (E–H) WGCNA gene dendrograms and module–trait correlation heatmaps for GSE135251 and the supportive RA peripheral-blood dataset GSE117769. (I) Sequential workflow in which 1299 and 1071 hepatic DEGs yielded 229 MASLD consensus genes, followed by intersection with 417 DEGs from GSE55457 and direction-concordance filtering to identify six shared genes. GSE77298 was not included in the primary 229-to-8-to-6 intersection. (J) Four-set Venn diagram showing the overall DEG overlap across the two hepatic and two RA synovial datasets. (K) Log2 fold-change patterns of the six shared genes across the four datasets. (L) GO biological-process enrichment of the six shared genes. (M) Expression of the six shared genes in NASH and control liver samples from GSE135251; Wilcoxon rank-sum test. (N) Pairwise Pearson correlations among the six shared genes in GSE135251. (O) WGCNA module–trait associations of the six shared genes in GSE135251. Boxplots show medians and interquartile ranges with individual observations.*P < 0.05, **P < 0.01, ***P < 0.001; ns, not significant. Abbreviations: DEG, differentially expressed gene; FC, fold change; GO, Gene Ontology; MASLD, metabolic dysfunction-associated steatotic liver disease; NAS, NAFLD activity score; NASH, non-alcoholic steatohepatitis; RA, rheumatoid arthritis; WGCNA, weighted gene co-expression network analysis. WGCNA was performed independently in GSE135251 and the supportive RA peripheral-blood dataset GSE117769. Fourteen co-expression modules were identified in GSE135251. MEtan and MEmagenta were positively correlated with fibrosis stage (r = 0.59 and r = 0.56, respectively), whereas MEblack was negatively correlated with MASLD status (r = −0.63) and NAS (r = −0.45). MEyellow was positively correlated with both fibrosis stage and NAS (Figure 4E and G). Eight modules were identified in GSE117769, with MEbrown and MEturquoise showing modest positive correlations with RA status (Figure 4F and H). The GSE117769 network was used as supportive co-expression evidence and did not contribute genes to the primary shared-gene intersection. Intersection of the 1299 and 1071 unique DEG sets from GSE135251 and GSE126848 yielded a 229-gene MASLD consensus set. Further intersection with the 417 RA synovial DEGs from GSE55457 produced eight candidate cross-disease genes. Requiring concordant log2 fold-change directions across GSE135251, GSE126848, and GSE55457 retained six shared genes: IL32, DHRS9, MOXD1, EGR1, RRM2, and PRC1 (Figure 4I–K). GSE77298 was not included in this primary 229-to-8-to-6 intersection and was reserved for independent RA synovial confirmation. In GSE135251, EGR1 was significantly downregulated, whereas IL32, DHRS9, RRM2, and PRC1 were significantly upregulated in NASH liver samples compared with healthy controls; MOXD1 did not differ significantly between the groups (P = 0.517; Figure 4M). Within the GSE135251 co-expression network, IL32 and MOXD1 were assigned to MEyellow, which was positively correlated with fibrosis stage and NAS, whereas EGR1 was assigned to MEblack, which was negatively correlated with MASLD status and NAS (Figure 4O). Pairwise Pearson analysis showed positive correlations among MOXD1, IL32, RRM2, and DHRS9 (r = 0.28–0.58), while EGR1 showed weaker correlations with the other shared genes (Figure 4N). GO-BP analysis of the six-gene set identified terms related to immune activation, cell-cycle regulation, and metabolic and proteolytic processes (Figure 4L). Machine-Learning Prioritisation Identified EGR1 as a Candidate Shared Gene Three machine-learning algorithms were applied to the 229-gene MASLD consensus set in GSE135251, comprising 206 MASLD samples and 10 healthy controls. LASSO regression retained FOS and CEBPD at λ.min, Random Forest ranked CEBPD highest by mean decrease in accuracy, with EGR1 among the top-ranked candidates, and SVM-RFE identified an optimal 32-gene feature subset (Figure 5A–C). FOS and CEBPD were common to the three algorithms but did not belong to the six directionally concordant cross-disease genes. EGR1 was supported by both Random Forest and SVM-RFE and was the only member of the six-gene shared set selected by at least two algorithms; it was therefore prioritised for subsequent analyses (Figure 5D). Within GSE135251, EGR1 showed an apparent AUC of 0.982 (95% CI, 0.964–0.999) for distinguishing MASLD from controls and was expressed at lower levels in MASLD samples (Wilcoxon P = 2.8 × 10−7; Figure 5E and F). Because candidate selection and performance assessment were conducted in the same highly imbalanced dataset, this AUC represents apparent rather than independently validated discriminatory performance. Machine-learning prioritisation and transcriptomic characterisation of EGR1 in MASLD. (A) LASSO coefficient paths and 10-fold cross-validated AUC. (B) Random Forest variable-importance ranking based on mean decrease in accuracy. (C) SVM-RFE feature-selection curve showing the optimal 32-gene subset. (D) Overlap of features selected by the three algorithms; EGR1 was the only shared gene supported by at least two algorithms. (E) Apparent ROC curve for EGR1 in GSE135251. (F) EGR1 expression in controls (n = 10) and MASLD samples (n = 206); Wilcoxon rank-sum test. (G) GSEA of the ranked GSE135251 transcriptome; positive and negative NES values indicate enrichment in NASH and controls, respectively. (H) GO biological-process enrichment of the 229 MASLD consensus genes. (I) Exploratory KEGG enrichment; pathways with nominal P < 0.05 were ranked by nominal P, and the top 20 are shown. (J) Curated EGR1 regulatory network based on TRRUST v2 and published literature. (K) Spearman co-expression network between EGR1 and the remaining MASLD consensus genes in GSE135251; genes with |ρ| > 0.4 and nominal P < 0.05 are shown. (L) Stage-stratified expression of the six shared genes across controls, NAFL, and NASH fibrosis stages F0–F4. In panel F, boxes indicate interquartile ranges and horizontal lines indicate medians. Abbreviations: AUC, area under the curve; GO, Gene Ontology; GSEA, gene set enrichment analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; LASSO, least absolute shrinkage and selection operator; MASLD, metabolic dysfunction-associated steatotic liver disease; NAFL, non-alcoholic fatty liver; NASH, non-alcoholic steatohepatitis; NES, normalised enrichment score; ROC, receiver operating characteristic; SVM-RFE, support vector machine with recursive feature elimination. GSEA of the 14,451-gene DESeq2 Wald-statistic ranking evaluated 19 predefined Hallmark-related gene sets, of which 11 met the Benjamini–Hochberg-adjusted P < 0.05 threshold (Figure 5G; Table S4). Gene sets related to cell-cycle progression, glycolysis, cholesterol homeostasis, and mTORC1 signalling were enriched in NASH, whereas TNFα/NF-κB signalling, inflammatory response, hypoxia, and oxidative phosphorylation were enriched in controls. GO-BP enrichment of the 229 MASLD consensus genes highlighted lipid- and steroid-metabolic processes and other hepatic biological functions (Figure 5H). In the exploratory KEGG analysis, pathways with nominal P < 0.05 were ranked by nominal P, and the top 20 were displayed; no pathway remained significant after Benjamini–Hochberg correction (Figure 5I). A TRRUST v2-derived regulatory network included reported upstream regulators of EGR1, including SP1, YAP1, JUN, and STAT3, and downstream targets related to extracellular-matrix regulation, inflammation, and cell-cycle control, including TGFB1, FN1, COL1A1, TNF, IL2, ICAM1, and RRM2 (Figure 5J). In the GSE135251 co-expression analysis, EGR1 showed its strongest positive correlations with FOS (ρ = 0.77) and JUN (ρ = 0.57), together with negative correlations with several genes involved in bile-acid transport and mitochondrial fatty-acid oxidation (Figure 5K). Stage-stratified profiling across the GSE135251 histological spectrum showed distinct expression patterns, with lower EGR1 and comparatively higher IL32, RRM2, PRC1, DHRS9, and MOXD1 expression in the more advanced disease groups (Figure 5L). External Validation of EGR1 Expression and Exploratory Diagnostic Modelling EGR1 expression was evaluated in three independent external validation cohorts. EGR1 was significantly downregulated in NASH hepatic tissue in GSE89632 (Wilcoxon P < 0.001) and in RA peripheral blood in GSE93272 (P < 0.001), with a directionally consistent but non-significant reduction in the smaller NAFLD/NASH cohort GSE48452 (P = 0.185, Figure 6A). External validation of EGR1 expression and exploratory diagnostic modelling. (A) EGR1 expression in the external hepatic cohorts GSE89632 and GSE48452 and the RA peripheral-blood cohort GSE93272; Wilcoxon rank-sum test. (B) EGR1 expression across controls, NAFL, and NASH fibrosis stages F0–F4 in GSE135251; Kruskal–Wallis test. (C) Stage-stratified EGR1 expression and comparison between mild (F0–F1) and advanced (F2–F4) fibrosis among patients with NASH; Wilcoxon rank-sum test for the two-group comparison. (D) ROC curves for EGR1 in GSE135251 and the external cohorts GSE89632 and GSE93272. (E) Apparent AUCs of the six shared genes and the combined six-gene model in GSE135251. (F) AUCs of EGR1 and the six shared genes for distinguishing mild from advanced fibrosis among patients with NASH. (G) Exploratory three-gene nomogram incorporating EGR1, IL32, and RRM2. (H) External expression replication of the six shared genes in GSE89632; Wilcoxon rank-sum test. (I) Bootstrap calibration of the three-gene model using 200 resamples. Boxplots show medians and interquartile ranges with individual observations. Abbreviations: AUC, area under the curve; CI, confidence interval; MASLD, metabolic dysfunction-associated steatotic liver disease; NAFL, non-alcoholic fatty liver; NASH, non-alcoholic steatohepatitis; RA, rheumatoid arthritis; ROC, receiver operating characteristic. Within GSE135251, EGR1 expression differed significantly across controls, NAFL, and NASH fibrosis stages F0–F4 and was generally lower in the more advanced disease groups (Kruskal–Wallis P = 8.99 × 10−8; Figure 6B). Among patients with NASH, EGR1 expression was significantly lower in advanced fibrosis (F2–F4) than in mild fibrosis (F0–F1) (Wilcoxon P = 0.001; Figure 6C). EGR1 showed an apparent AUC of 0.982 (95% CI, 0.964–0.999) for distinguishing MASLD from controls in GSE135251 and an AUC of 0.989 (95% CI, 0.966–1.000) for distinguishing NASH from controls in the external hepatic cohort GSE89632. Its AUC for distinguishing RA from controls in the peripheral-blood cohort GSE93272 was 0.668 (95% CI, 0.587–0.748) (Figure 6D). Among the six shared genes evaluated individually in GSE135251, EGR1 had the highest apparent AUC (0.982), followed by PRC1 (0.844), IL32 (0.793), DHRS9 (0.778), RRM2 (0.720), and MOXD1 (0.439). A combined six-gene logistic-regression model yielded an apparent AUC of 0.993 (95% CI, 0.984–1.000) (Figure 6E). In the fibrosis-staging analysis comparing F0–F1 with F2–F4 among patients with NASH, MOXD1 had the highest AUC (0.828), followed by IL32 (0.760) and EGR1 (0.682) (Figure 6F). An exploratory multivariable logistic-regression nomogram incorporating EGR1, IL32, and RRM2 was constructed to visualise predicted MASLD probability, with EGR1 contributing the largest point range (Figure 6G). Bootstrap calibration with 200 resamples showed close apparent agreement between predicted and observed probabilities, with a Hosmer–Lemeshow goodness-of-fit P value of 1.000 (Figure 6I). External expression analysis of the six shared genes in GSE89632 supported differential expression in the expected directions for EGR1, DHRS9, IL32, and PRC1 (all P ≤ 0.025), whereas RRM2 and MOXD1 did not differ significantly between NASH and controls (Figure 6H). Because the single-gene, six-gene, and three-gene models in GSE135251 were fitted and evaluated in the same highly imbalanced dataset, their performance estimates should be interpreted as apparent or internally calibrated rather than independently validated. Macrophage-Related Immune Signatures in MASLD and Cell-Type Localisation of Shared Genes in RA Synovium Immune-signature enrichment in GSE135251 was assessed by ssGSEA using nine predefined custom marker-gene sets and examined across fibrosis stages (Figure 7A). M1 macrophage scores differed significantly across fibrosis stages (Kruskal–Wallis P = 0.00013; Figure 7B), while M2 macrophage scores also showed stage-dependent differences (Figure 7G). Among the nine immune-cell signatures, the M1 macrophage score showed the strongest positive association with ordinal fibrosis stage (Spearman r = 0.32; Figure 7I). Macrophage-related immune signatures in MASLD and cell-type localisation of shared genes in RA synovium. (A) Stage-level heatmap of standardised ssGSEA scores for nine immune-cell signatures across MASLD fibrosis stages. (B) M1 macrophage signature scores across fibrosis stages; Kruskal–Wallis test. (C) Pearson correlation between the six-gene module score and the M1 macrophage signature score. (D) Pearson correlation between EGR1 expression and the M2 macrophage signature score. (E) Immune-signature scores in EGR1-high and EGR1-low groups; Wilcoxon rank-sum test. (F) Pearson correlations between the six shared genes and nine immune-cell signature scores. (G) M1 and M2 macrophage signature scores across fibrosis stages. (H) M1 and M2 macrophage signature scores according to EGR1 expression group; Wilcoxon rank-sum test. (I) Spearman correlations between immune-cell signature scores and ordinal fibrosis stage. (J) UMAP representation of 38,259 RA synovial cells from GSE200815, consolidated into six major cell populations. (K) EGR1 expression across the annotated RA synovial cell populations. (L) Expression and detection frequency of the six shared genes across RA synovial cell populations. Boxplots show medians and interquartile ranges with individual observations. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001; ns, not significant. Abbreviations: MASLD, metabolic dysfunction-associated steatotic liver disease; RA, rheumatoid arthritis; ssGSEA, single-sample gene set enrichment analysis; UMAP, uniform manifold approximation and projection. The composite six-gene module score was positively correlated with the M1 macrophage score (Pearson r = 0.464, P = 6.53 × 10−13; Figure 7C), with DHRS9 and PRC1 showing the strongest positive gene-level correlations with the M1 signature (Figure 7F). EGR1 expression was negatively correlated with the M2 macrophage score (Pearson r = −0.284, P = 2.33 × 10−5; Figure 7D). When samples were dichotomised at the median EGR1 expression level, the EGR1-high group had a lower M2 macrophage score, whereas the M1 macrophage score did not differ significantly between groups (Figure 7H). Differences were also observed in B-cell, dendritic-cell, neutrophil, and natural-killer-cell signature scores between the EGR1-high and EGR1-low groups (Figure 7E). Single-cell RNA sequencing of RA synovium from GSE200815 was used to examine the cellular distribution of the six shared genes. After quality control and Harmony integration, 38,259 cells were consolidated into six major populations: fibroblasts, endothelial cells, T cells, macrophages, B cells, and plasma cells (Figure 7J). EGR1 was detected across multiple synovial cell populations, with comparatively higher expression in plasma cells and fibroblasts (Figure 7K). Among the six shared genes, IL32 was most prominently expressed in T cells, whereas EGR1 showed relatively higher expression in fibroblasts and plasma cells (Figure 7L). Experimental Validation Supported a Coordinated EGR1–IL-32–Macrophage Polarisation Signature Serum IL-32 concentrations increased across the HC, RA-only, and RA+MASLD groups (89.5 ± 40.2, 237.2 ± 49.7, and 401.4 ± 54.1 pg/mL, respectively; Kruskal–Wallis P < 0.001; Figure 8A), with all pairwise comparisons remaining significant after Bonferroni correction. Conversely, PBMC EGR1 mRNA expression was lower in both RA groups than in HC (1.416 ± 0.493, 0.889 ± 0.352, and 0.631 ± 0.309, respectively; Kruskal–Wallis P < 0.001). After Bonferroni correction, the differences were significant between HC and RA only (P = 0.003) and between HC and RA+MASLD (P < 0.001), whereas the difference between the two RA groups was not significant (P = 0.267; Figure 8B). In the exploratory subset of three randomly selected participants per group, Western blotting similarly showed lower EGR1 protein abundance in the RA groups, with the lowest level in RA+MASLD (Figure 8C and D). Patient-level expression and exploratory cellular assessment of the EGR1–IL-32–macrophage polarisation signature. (A) Serum IL-32 concentrations in healthy controls (HC), patients with RA only, and patients with RA+MASLD (n = 15 per group). (B) Relative EGR1 mRNA expression in peripheral blood mononuclear cells (n = 15 per group). (C) Representative Western blots of EGR1 and GAPDH in the randomly selected exploratory subset (n = 3 per group). (D) Densitometric quantification of EGR1 protein expression normalised to GAPDH. (E) Representative CD68/CD86 and CD68/CD163 flow-cytometric plots within the CD45⁺ leukocyte gate; the complete gating strategy is shown in Figure S2. (F) M1-like/M2-like ratio. (G) Frequency of M1-like cells, defined as CD45⁺CD68⁺CD86⁺ cells and expressed as a percentage of total CD45⁺ cells. (H) Frequency of M2-like cells, defined as CD45⁺CD68⁺CD163⁺ cells and expressed as a percentage of total CD45⁺ cells. Flow-cytometric analyses were exploratory and included three randomly selected participants per group. (I) Spearman correlation heatmap of EGR1 mRNA and serum IL-32 with clinical measures; the EGR1–IL-32 correlation was assessed in all 45 participants, whereas clinical correlations were assessed in patients with RA (n = 30). Bars show means with standard deviations. Overall group differences were assessed using the Kruskal–Wallis test or one-way ANOVA, as appropriate, with multiplicity-adjusted pairwise comparisons. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001; ns, not significant. Abbreviations: ALT, alanine aminotransferase; BMI, body mass index; CAP, controlled attenuation parameter; CRP, C-reactive protein; DAS28, 28-joint Disease Activity Score; FIB-4, fibrosis-4 index; GAPDH, glyceraldehyde-3-phosphate dehydrogenase; HC, healthy control; LSM, liver stiffness measurement; MASLD, metabolic dysfunction-associated steatotic liver disease; PBMC, peripheral blood mononuclear cell; RA, rheumatoid arthritis. Flow-cytometric analysis in the same exploratory subset showed reciprocal differences in M1-like and M2-like cell frequencies (Figure 8E; Figure S2). M1-like cells, defined as CD45⁺CD68⁺CD86⁺ cells and expressed as a percentage of total CD45⁺ cells, increased across HC, RA only, and RA+MASLD (3.24 ± 0.73%, 6.73 ± 0.41%, and 9.62 ± 0.33%, respectively; one-way ANOVA P = 1.75 × 10−5; Figure 8G). In contrast, M2-like cells, defined as CD45⁺CD68⁺CD163⁺ cells and expressed using the same denominator, decreased across the three groups (8.12 ± 0.15%, 5.89 ± 0.41%, and 3.38 ± 0.38%, respectively; one-way ANOVA P = 7.26 × 10−6; Figure 8H). The M1-like/M2-like ratio correspondingly increased from 0.40 ± 0.08 in HC to 1.15 ± 0.15 in RA only and 2.87 ± 0.34 in RA+MASLD (Kruskal–Wallis P = 0.027; Figure 8F). These exploratory cellular findings were directionally consistent with the macrophage-related immune-signature patterns observed in GSE135251. Across all 45 participants, PBMC EGR1 mRNA expression was inversely correlated with serum IL-32 concentration (Spearman ρ = −0.663, P < 0.001; Figure 8I). Among the 30 participants with RA, serum IL-32 was positively correlated with multiple inflammatory, metabolic, and hepatic measures (ρ = 0.381–0.817), whereas EGR1 mRNA was inversely correlated with several corresponding measures (ρ = −0.381 to −0.608; all P < 0.05). Overall, these findings supported coordinated changes in EGR1, IL-32, and peripheral M1-like/M2-like phenotypes, although the Western blot and flow-cytometric results remained exploratory because they were based on three participants per group. Discussion In this multi-tier study, RA was not independently associated with MASLD or the assessed hepatic phenotypes after full cardiometabolic adjustment in NHANES, whereas patients with RA+MASLD in the hospital cohort had greater inflammatory and metabolic burden, hepatocellular injury, and noninvasive fibrosis burden than those with RA alone. Cross-disease transcriptomic integration identified six concordantly dysregulated genes and prioritised EGR1 for subsequent analyses, while immune-signature and single-cell analyses provided macrophage-related and cell-type context. At the patient level, PBMC EGR1 mRNA showed a stepwise decrease and serum IL-32 progressively increased across the healthy-control, RA-only, and RA+MASLD groups; exploratory Western blot and flow-cytometric findings in three participants per group were directionally consistent with these observations. Collectively, these findings support a coordinated EGR1–IL-32–macrophage polarisation signature associated with RA–MASLD comorbidity, but do not establish causal directionality or a confirmed signalling pathway. The differing findings from NHANES and the hospital cohort should be interpreted in the context of their distinct study designs and populations. In NHANES, RA was not independently associated with MASLD or the assessed hepatic phenotypes after full adjustment, consistent with the non-significant adjusted estimates reported in a previous US population-based analysis,11 whereas an independent RA–MASLD association was observed in the European Paracelsus 10,000 cohort.12 Several factors may have contributed to this heterogeneity, including differences in RA ascertainment—self-reported physician-diagnosed RA in NHANES versus classification according to the 2010 ACR/EULAR criteria in the hospital cohort28,29—the high background prevalence of MASLD in the general population,30 and the inability of cross-sectional data to capture disease activity, treatment history, or cumulative inflammatory exposure over time. By contrast, the hospital cohort comprised clinically classified patients with RA; those with RA+MASLD had higher disease activity, metabolic burden, and noninvasive fibrosis indices, while BMI, triglycerides, and DAS28 retained associations with MASLD in the fitted multivariable model. These findings are broadly consistent with hepatic abnormalities reported in other clinically characterised RA cohorts,31 although the selected single-centre setting and differences in medication exposure may also have contributed to the observed contrasts. These differences may partly explain why the clinical and molecular signals were more apparent in the selected hospital cohort than in the population-based analysis. Similar heterogeneity between population-based and clinically selected cohorts has been described at other immune-mediated–metabolic interfaces, including psoriatic disease and NAFLD.32 EGR1 was prioritised for subsequent analyses because it was supported by both Random Forest and SVM-RFE and was the only member of the six directionally concordant shared genes selected by at least two algorithms. The reduced EGR1 expression observed in MASLD liver and RA peripheral blood should not be interpreted as evidence of a uniformly pro- or anti-fibrotic function. Previous studies have implicated EGR1 in fibrogenic signalling and tissue remodelling.33,34 More specifically, EGR1 has been linked to TGF-β1/Smad3/BRD4-associated activation in hepatic stellate cells, and Egr1 knockdown attenuated fibrotic changes in a thioacetamide-induced liver-injury model.35 Conversely, Egr1 deficiency aggravated acute carbon tetrachloride-induced liver injury and delayed hepatocyte cell-cycle entry and regeneration.36,37 These apparently divergent findings suggest that the effects of EGR1 may depend on cell type, injury context, and disease stage, as well as on the balance between fibrogenic signalling and tissue repair. In human monocyte-to-macrophage differentiation, Trizzino et al demonstrated that EGR1 restrains inflammatory enhancer activity, providing a plausible context in which reduced EGR1 expression could accompany enhanced inflammatory activity.23 Our single-cell analysis detected EGR1 across several RA synovial cell populations, with comparatively higher expression in fibroblasts and plasma cells; however, these data do not identify the cellular source of the reduced PBMC EGR1 signal. The TRRUST-derived network and EGR1 co-expression patterns should therefore be regarded as hypothesis-generating pending functional validation.38 Serum IL-32 showed marked between-group separation, increasing approximately 4.5-fold from healthy controls to patients with RA+MASLD, whereas PBMC EGR1 mRNA changed in the opposite direction and was inversely correlated with serum IL-32 across all 45 participants. Although this reciprocal pattern is consistent with coordinated changes in the two markers, the cross-sectional association does not establish direct regulation or temporal order. IL-32 has been implicated in RA synovial inflammation and evaluated as a potential biomarker in rheumatic diseases,39,40 while IL-32γ has been reported as a highly active isoform.41 Saturated fatty acids can induce IL-32 expression in hepatocytes,42 providing a potential metabolic context for its elevation in MASLD; however, whether this mechanism contributes to circulating IL-32 in RA–MASLD comorbidity remains unknown. Among participants with RA, serum IL-32 was positively associated with disease activity and several metabolic and hepatic measures, whereas EGR1 mRNA showed inverse associations with a subset of these variables. Because these clinical measures are interrelated and were assessed cross-sectionally, the observed correlations should not be interpreted as evidence that EGR1 directly regulates IL-32 or that either marker independently drives hepatic disease. Moreover, the ELISA quantified total circulating IL-32 rather than individual isoforms, and the cellular source of serum IL-32 was not determined. These findings therefore support a coordinated EGR1–IL-32 signature associated with clinical disease burden, while requiring isoform-specific and functional validation. The macrophage-related findings showed directional concordance across analytical layers, despite arising from distinct tissues and measurement frameworks. In GSE135251, the M1 macrophage signature score showed the strongest positive association with ordinal fibrosis stage among the evaluated immune signatures. This finding is consistent with the M1-skewed macrophage phenotype reported in RA synovial inflammation43,44 and with the recognised involvement of pro-inflammatory hepatic macrophages in lipotoxic injury and fibrosis progression in MASLD.45,46 In the exploratory flow-cytometric subset, M1-like cell frequencies increased, whereas M2-like cell frequencies decreased, from healthy controls to the RA-only and RA+MASLD groups. EGR1 expression was inversely associated with the M2 macrophage signature score, consistent with previous evidence linking EGR1 to transcriptional programmes involved in macrophage polarisation and proliferation.47 Together, these observations support a macrophage-related component of the coordinated EGR1–IL-32 signature and provide a rationale for further functional investigation. Because ssGSEA and peripheral-blood flow cytometry capture different biological compartments, confirmation in larger cohorts with tissue-specific analyses is warranted, particularly given that the flow-cytometric assessment included only three participants per group. This study integrated population-level, clinical, transcriptomic, single-cell, and patient-level evidence, but several limitations define the scope of interpretation. RA ascertainment in NHANES relied on self-reported physician diagnosis, which may have introduced misclassification, and its cross-sectional design precluded assessment of temporal relationships.48,49 Hepatic phenotypes were evaluated using noninvasive measures rather than concurrent liver histology, and liver stiffness may be influenced by obesity and hepatic inflammation in addition to fibrosis.50 The hospital cohort was retrospective and single-centre, while differences in antirheumatic treatment exposure may have contributed to the observed clinical and experimental patterns.51 In addition, the transcriptomic analyses spanned different tissues, platforms, and disease settings, introducing biological and technical heterogeneity despite gene-level integration. The high AUCs obtained in GSE135251 represent apparent performance because feature prioritisation, model fitting, and evaluation were conducted in the same highly imbalanced dataset, and the multigene models were not externally validated. Finally, ssGSEA provided relative immune-signature enrichment scores, whereas Western blotting and flow cytometry were performed in only three participants per group and should therefore be considered exploratory. Future studies should include prospective multicentre and multi-ethnic cohorts, independent validation of the predictive models, and liver histology where clinically appropriate. Functional perturbation of EGR1 and IL-32 in relevant immune and hepatic cell populations, together with genetic causal analyses using EGR1 cis-eQTLs, may help clarify regulatory direction and biological function. It also remains to be determined whether emerging MASH therapies that improve steatohepatitis or fibrosis modify the EGR1–IL-32–macrophage signature in patients with RA.52,53 Conclusions In this multi-tier study, RA was not independently associated with MASLD or the assessed hepatic phenotypes after full cardiometabolic adjustment in NHANES, whereas the hospital cohort showed greater inflammatory, metabolic, hepatic, and fibrosis-related burden in patients with RA+MASLD. Cross-disease transcriptomic, immune-signature, single-cell, and patient-level analyses converged on a coordinated EGR1–IL-32–macrophage polarisation signature associated with RA–MASLD comorbidity. These observational findings require prospective multicentre replication and functional validation before the signature can be considered for biomarker development or therapeutic application. Acknowledgments We thank the staff of the Department of Rheumatology and the clinical laboratory of the First Affiliated Hospital of Anhui University of Chinese Medicine for their support, and the participants who consented to provide data and samples. We also acknowledge the public GEO and NHANES resources that made this study possible. Funding Statement This work was supported by the National Natural Science Foundation of China General Program (82574970); the 2025 Anhui Provincial Department of Education Scientific Research Project for Young Scholars (2025AHGXZK40325); the Major and Difficult Diseases Clinical Research Project on Collaboration between Traditional Chinese and Western Medicine (ZDYN-2024-A-146); the Evidence-Based Research System for Traditional Chinese Medicine project and the Traditional Chinese Medicine Standards Research and Promotion Base Construction project (both under Anhui Finance and Society2025 No. 1382); and the Institute of Xin’an Medicine and Modernization of Traditional Chinese Medicine, Great Health Research Institute Special Projects (2023CXMMTCM004). The funders had no role in study design, data collection, data analysis, data interpretation, or writing of the report. Use of Artificial Intelligence-Assisted Technology During preparation of this paper, the authors used ChatGPT (OpenAI) for language editing, formatting assistance, and checking compliance with journal formatting requirements. No artificial intelligence tool was used to generate or analyse data, interpret results, or draw scientific conclusions. The authors reviewed and edited all outputs from this assistance and take full responsibility for the final content. Abbreviations ACR/EULAR, American College of Rheumatology/European League Against Rheumatism; ALP, alkaline phosphatase; ALT, alanine aminotransferase; APRI, AST-to-platelet ratio index; AST, aspartate aminotransferase; AUC, area under the curve; b/tsDMARD, biological or targeted synthetic disease-modifying antirheumatic drug; BMI, body mass index; CAP, controlled attenuation parameter; CI, confidence interval; CRP, C-reactive protein; DAS28, 28-joint Disease Activity Score; DEG, differentially expressed gene; DMARD, disease-modifying antirheumatic drug; ELISA, enzyme-linked immunosorbent assay; ESR, erythrocyte sedimentation rate; FIB-4, fibrosis-4 index; GEO, Gene Expression Omnibus; GO-BP, Gene Ontology biological process; GSEA, gene set enrichment analysis; HC, healthy control; HDL, high-density lipoprotein; IL-32, interleukin-32; KEGG, Kyoto Encyclopedia of Genes and Genomes; LASSO, least absolute shrinkage and selection operator; LSM, liver stiffness measurement; MASLD, metabolic dysfunction-associated steatotic liver disease; MASH, metabolic dysfunction-associated steatohepatitis; NAFLD, non-alcoholic fatty liver disease; NASH, non-alcoholic steatohepatitis; NHANES, National Health and Nutrition Examination Survey; OR, odds ratio; PBMC, peripheral blood mononuclear cell; qRT-PCR, quantitative reverse-transcription polymerase chain reaction; RA, rheumatoid arthritis; ROC, receiver operating characteristic; RNA-seq, RNA sequencing; ssGSEA, single-sample gene set enrichment analysis; SVM-RFE, support vector machine with recursive feature elimination; UMAP, uniform manifold approximation and projection; WGCNA, weighted gene co-expression network analysis. Data Sharing Statement All GEO datasets analysed in this study are publicly available under the accession numbers listed in the Methods section (GSE135251, GSE126848, GSE77298, GSE55457, GSE89632, GSE48452, GSE93272, GSE117769, and GSE200815). NHANES data are publicly accessible at https://www.cdc.gov/nchs/nhanes/. De-identified participant-level data from the hospital-based cohorts and the analytic R scripts are available from the corresponding author on reasonable request, subject to institutional approval. Ethics Approval and Informed Consent The NHANES 2017–March 2020 protocol was approved by the National Center for Health Statistics Research Ethics Review Board, and written informed consent was obtained from all NHANES participants. The hospital-based study was approved by the Ethics Committee of the First Affiliated Hospital of Anhui University of Chinese Medicine (Approval No. 2023AH-76). For the retrospective clinical cohort (n = 100), which relied on existing clinical, laboratory, and imaging records, the requirement for individual informed consent was waived by the Ethics Committee. For the prospective experimental validation sub-study (n = 45; ELISA, qRT-PCR, Western blot, and flow cytometry), which involved fresh biological sample collection, written informed consent was obtained from every participant prior to sample collection under the same approval. All procedures were conducted in accordance with the Declaration of Helsinki. Author Contributions LP: Conceptualization, Formal analysis, Data curation, Investigation, and Writing – original draft. YL: Data curation, Investigation, and Writing – review and editing. LC: Investigation and Writing – review and editing. CH: Conceptualization, Funding acquisition, Supervision, and Writing – review and editing. LP and CH verified the underlying data and accept responsibility for the decision to submit the paper for publication. All authors gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work. Disclosure The authors declare support for the paper from National Natural Science Foundation of China General Program (No. 82574970); the 2025 Anhui Provincial Department of Education Scientific Research Project for Young Scholars (No. 2025AHGXZK40325); Major and Difficult Diseases Clinical Research Project on Collaboration between Traditional Chinese and Western Medicine (No. ZDYN-2024-A-146); Evidence-Based Research System for Traditional Chinese Medicine project and Traditional Chinese Medicine Standards Research and Promotion Base Construction project (Anhui Finance and Society [2025] No. 1382); Institute of Xin’an Medicine and Modernization of Traditional Chinese Medicine, Great Health Research Institute Special Projects (No. 2023CXMMTCM004). The authors declare that they have no other competing interests. References Associated Data Data Availability Statement All GEO datasets analysed in this study are publicly available under the accession numbers listed in the Methods section (GSE135251, GSE126848, GSE77298, GSE55457, GSE89632, GSE48452, GSE93272, GSE117769, and GSE200815). NHANES data are publicly accessible at https://www.cdc.gov/nchs/nhanes/. De-identified participant-level data from the hospital-based cohorts and the analytic R scripts are available from the corresponding author on reasonable request, subject to institutional approval.