Clinical and Experimental Hepatology

Ferritin-derived composite indices are associated with MAFLD and support interpretable risk prediction: evidence from NHANES 2017–2023 and NHANES III

  1. Department of Laboratory Medicine, Xinyu Iron and Steel Group Co. Ltd., Central Hospital, China

  2. Center of Hepato-Pancreato-Biliary Surgery, Department of Pancreato-Biliary Surgery, The First Affiliated Hospital of Sun Yat-sen University, China

  3. Department of Laboratory Medicine, The First Affiliated Hospital of Sun Yat-sen University, China

  4. Shenzhen Hospital, Southern Medical University (Shenzhen School of Clinical Medicine, Southern Medical University), China

Clin Exp HEPATOL 2026; 12, 4:

Data publikacji online: 2026/11/19
Article files
Ferritin Wang 00839.pdf Ferritin Wang 00839 supl.pdf

Introduction

Metabolic dysfunction-associated fatty liver disease (MAFLD) ranks among the most common chronic liver conditions globally. Its prevalence has risen from 17.6% in 1990 to an estimated 33.6% in 2025, representing about 1.27 billion cases. Asia and Latin America bear the highest burden (> 44%) [1-3]. The disease spans simple steatosis, metabolic dysfunction-associated steatohepatitis (MASH), cirrhosis, and hepatocellular carcinoma. Beyond the liver, MAFLD independently raises the risks of cardiovascular disease, chronic kidney disease, and several extrahepatic cancers [4-6]. Despite its high frequency, MAFLD often goes undiagnosed because current detection relies on traditional metabolic factors – obesity, diabetes, and dyslipidemia – that do not fully reflect the disease’s complex biology [7]. The clinical variability and weak correlation between conventional markers and steatosis severity highlight an urgent need for accessible, biologically meaningful biomarkers to improve risk stratification and early detection.

Iron metabolism has long been connected to metabolic and liver disorders, but its relation to MAFLD is intricate [8-10]. Ferritin acts both as an iron‑storage protein and as an acute-phase reactant that rises with chronic inflammation. Hyperferritinemia occurs in about 25% of MAFLD patients, with the highest rates in Asian populations, and it identifies individuals with more severe metabolic dysfunction, raised liver enzymes, insulin resistance, and higher liver-related event risk [11]. NHANES-based studies have confirmed ferritin’s association with MAFLD and fibrosis, and Mendelian randomization supports a causal contribution beyond a mere epiphenomenon [12-14]. Mechanistically, excess iron can trigger lipid peroxidation, mitochondrial damage, and ferroptosis, accelerating progression to steatohepatitis and fibrosis [15]. Yet ferritin levels are affected by nutritional status, systemic inflammation, and hepatic synthetic function – factors that vary widely [16]. Thus, raw ferritin alone gives an incomplete measure of the biologically relevant iron burden in MAFLD.

Earlier work has treated ferritin as a standalone marker or, less commonly, used traditional iron panel parameters. To our knowledge, no study has systematically compared ferritin-derived composite indices that combine ferritin with albumin, lymphocyte count, or hemoglobin – e.g., ferritin-to-albumin ratio (FAR), ferritin-to-lymphocyte ratio (FLR), ferritin-to-hemoglobin ratio (FHR), and ferritin-to-albumin-to-lymphocyte ratio (FALR) – to capture nutritional, inflammatory, and hematological dimensions at once [17-20]. Also, no investigation has evaluated their dose-response shapes, threshold effects, and subgroup consistency in a nationally representative sample, nor integrated them into interpretable machine-learning models with external validation in an independent historical cohort. The lack of multi-dataset evaluation and external replication limits generalizability and impedes translation into clinical practice.

We therefore used NHANES 2017–2023 as the development cohort and NHANES III for external validation. Our objectives were: 1) to assess cross-sectional associations of FALR, FAR, FLR, and FHR with MAFLD using multivariable logistic regression; 2) to characterize dose-response relationships and potential threshold effects via restricted cubic splines and piecewise models; 3) to test subgroup consistency; 4) to develop and externally validate an interpretable GBM model incorporating the most robust indices; and 5) to apply SHAP and LIME for model transparency. By drawing on two large U.S. cohorts separated by three decades, we aimed to provide replicable evidence and an externally validated, clinically accessible tool for risk identification.

Material and methods

Study design, data sources, and study population

The National Health and Nutrition Examination Survey (NHANES) is a stratified, multistage probability survey of the civilian, noninstitutionalized U.S. population conducted by the National Center for Health Statistics (NCHS), part of the Centers for Disease Control and Prevention (CDC), using household interviews, standardized examinations, and cycle-specific laboratory protocols; NHANES III followed the same general design [21-23]. We used public-use files from NHANES 2017–2020, 2021–2023, and NHANES III, merged by respondent sequence number using official data dictionaries; no restricted-access or identifiable data were used. For the 2017–2023 development cohort (initial n = 27,493), we restricted the sample to adults ≥ 20 years and excluded those with missing controlled attenuation parameter (CAP; n = 3,505) and other required variables (n = 5,092), yielding 8,444 participants (3,512 MAFLD cases, 41.6%). For NHANES III external validation, we applied the same age restriction and excluded participants with missing data for MAFLD classification, retaining 12,918 individuals (2,608 MAFLD cases, 20.2%). Figure 1 details the selection process. All original NHANES data collections were approved by the NCHS Ethics Review Board with appropriate participant consent. Because this secondary analysis used de-identified public data only, with no participant contact or access to identifiable information, it did not constitute human subjects research under 45 CFR 46.102(e) and required no additional IRB review. The analysis complied with the Declaration of Helsinki and the NCHS Data User Agreement.

Exposure and outcome

Four ferritin-derived indices were calculated: FALR = ferritin/(albumin × lymphocyte); FAR = ferritin/albumin; FLR = ferritin/lymphocyte; and FHR = ferritin/hemoglobin (ferritin in ng/ml, albumin in g/dl, lymphocyte count in 103/µl, hemoglobin in g/dl). Each was analyzed continuously and in quartiles. In NHANES 2017–2023, MAFLD was defined as CAP ≥ 274 dB/m plus ≥ 1 of: overweight/obesity (body mass index [BMI] ≥ 25 [≥ 24 for Asians] or waist criteria), type 2 diabetes, or metabolic dysregulation (≥ 2 of elevated blood pressure, triglycerides, reduced high-density lipoprotein cholesterol [HDL-C], or prediabetes). In NHANES III, moderate or severe steatosis on ultrasound was considered positive; the same metabolic criteria were mapped to available variables, and missing data were excluded. Given the different steatosis assessment modalities (CAP vs. ultrasound), we view this as a temporal external validation under a harmonized but nonidentical outcome definition.

Covariates and imputation

Covariates included age, sex, race/ethnicity, education, and poverty income ratio (PIR) (Model 2), and additionally BMI, smoking, alcohol, hypertension, diabetes (Model 3). Missing covariates were imputed using random forest, including the outcome to inform imputation but never imputing the outcome or the four indices themselves. For machine learning, the development set was split 70:30; all preprocessing and tuning were performed within the training set using 5-fold CV; NHANES III was used only for final validation.

Statistical analysis

Weighted logistic regression estimated associations (Models 1-3), with quartile trend tests using median-assigned ordinal variables. Restricted cubic splines (4 knots) and likelihood-ratio tests assessed nonlinearity. Two‑piecewise threshold models identified inflection points (K) via grid search; log-likelihood ratio tests compared linear vs. threshold models. Subgroup analyses by sex, age, BMI, smoking, alcohol, diabetes, and hypertension used multiplicative interaction terms. Two-sided p < 0.05 was significant; analyses used R (survey, rms).

Machine learning

Thirty-four candidate variables were considered. Six feature-selection algorithms (LASSO, Boruta, Random Forest, XGBoost, SVM-RFE, Genetic Algorithm) were applied; their intersection yielded 12 predictors (FAR, FHR, race, albumin, hemoglobin, iron, total iron-binding capacity [TIBC], alkaline phosphatase [ALP], creatinine, total cholesterol, white blood cell count [WBC], red blood cell count [RBC]). We chose intersection to prioritize parsimony and reproducibility over maximal performance; individual algorithm outputs are in the supplement. Seven models were evaluated; GBM was selected by internal test AUC (0.731), refit on the full development cohort, and externally validated in NHANES III. Performance metrics included AUC, accuracy, sensitivity, specificity, F1. SHAP and LIME provided interpretability; a web-based prototype risk calculator was developed for research use.

Results

Baseline characteristics of the study populations

Figure 1 shows the study flow. From 27,493 participants initially screened in NHANES 2017–2023, we excluded those aged < 20 years (n = 10,452), those with missing CAP (n = 3,505), and those with missing required variables (n = 5,092). The final development cohort included 8,444 participants, with 3,512 MAFLD cases (41.6%). For external validation, we retained 12,918 ultrasound-evaluable adults from NHANES III, including 2,608 MAFLD cases (20.2%).

Table 1 and Supplementary Table 1 summarize baseline characteristics by MAFLD status. In the development cohort, MAFLD participants were older and had higher BMI. Serum ferritin was markedly higher in MAFLD cases, whereas albumin was slightly lower. All four composite indices – FALR, FAR, FLR, and FHR – were significantly higher in the MAFLD group. MAFLD cases also had a greater proportion of males, higher prevalences of hypertension and diabetes, and were more often former smokers. The NHANES III external cohort showed similar patterns, though effect sizes differed slightly between datasets.

Association between ferritin‑derived composite indices and MAFLD

Table 2 presents the associations from continuous and quartile-based logistic regression models with progressive adjustment. In Model 1 (unadjusted), all four indices showed significant positive associations in both continuous and categorical forms (all p < 0.05). After full adjustment (Model 3), the associations attenuated but remained significant for FAR and FHR (Fig. 2).

In the development cohort (Table 2), FAR gave the most robust association: the fully adjusted odds ratio (OR) for the highest vs. lowest quartile (Q4 vs. Q1) was 1.69. Continuous FAR also remained significant. FHR similarly showed a strong association (Q4 vs. Q1 OR = 1.65, 95% CI: 1.28-2.13; p for trend < 0.001; continuous OR = 1.012, 95% CI: 1.005-1.018, p = 0.001). FALR and FLR also had positive associations but with smaller effect sizes: FALR Q4 vs. Q1 OR = 1.33 with a borderline trend p (0.052), and FLR Q4 vs. Q1 OR = 1.38. The attenuation from Model 1 to Model 3 was greater for FALR and FLR than for FAR and FHR, suggesting that the latter two may capture MAFLD-related information less confounded by metabolic factors.

When we applied the same quartile cut‑points from the development cohort to NHANES III (Supplementary Table 2), all four indices showed substantially stronger associations. Fully adjusted Q4 vs. Q1 ORs ranged from 2.34 for FALR to 3.00 for FHR, with all p-values for trend < 0.001. Continuous associations were also significant across all models in NHANES III. These findings confirm the reproducibility of the positive directional association across independent historical cohorts, though differences in effect magnitudes likely reflect variations in measurement methods, case definitions, and population characteristics between eras.

Nonlinear and threshold effects

To explore dose-response shapes beyond linearity, we performed restricted cubic spline analyses (Fig. 3). In the development cohort, all four indices showed significant overall associations and nonlinear relationships with MAFLD risk (p for overall < 0.001, p for nonlinearity < 0.001). The RCS curves consistently showed steep increases at lower-to-moderate levels, with slopes flattening at higher levels – a saturation pattern.

Piecewise threshold models (Table 3, Figs. 4 and 5) quantified this biphasic pattern. For FAR, the inflection point was 60.42. Below this threshold, each one‑unit increase in FAR raised MAFLD risk by 1.1% (OR = 1.011, 95% CI: 1.008-1.014, p < 0.001); above it, the slope flattened (OR = 0.999, 95% CI: 0.998-1.001, p = 0.506). FHR had an inflection point at 15.10, with a sharp risk rise below (OR = 1.041, 95% CI: 1.028-1.054, p < 0.001) and a null effect above (OR = 0.999, 95% CI: 0.994-1.004, p = 0.747). FALR and FLR showed analogous patterns, with inflection points at 7.20 and 29.04, respectively, and significant risk increments below the thresholds (FALR: OR = 1.067, 95% CI: 1.035-1.101, p < 0.001; FLR: OR = 1.017, 95% CI: 1.009-1.024, p < 0.001), followed by nonsignificant slopes above. Log-likelihood ratio tests comparing the linear model with the two-piecewise model were significant for all four indices, confirming better fit for the threshold models.

Applying the same development-derived thresholds to NHANES III (Supplementary Table 3) largely replicated the pattern for FAR and FHR: both had significant positive slopes below the inflection points (FAR: OR = 1.020, p < 0.001; FHR: OR = 1.076, p < 0.001) and flat slopes above (FAR: OR = 1.001, p = 0.696; FHR: OR = 1.007, p = 0.207). For FALR and FLR, the slopes above thresholds, though attenuated, remained statistically significant in the external set, suggesting that the saturation effect may be less complete for these two indices in the historical cohort.

Subgroup analyses

We assessed consistency across subgroups: sex, age (< 60 vs. ≥ 60 years), BMI (< 30 vs. ≥ 30 kg/m2), smoking, alcohol, diabetes, and hypertension (Fig. 6; Supplementary Fig. 4 for NHANES III). In the development cohort, the positive associations (Q4 vs. Q1) were generally consistent across all strata. Although a few nominally significant interactions appeared (e.g., slightly stronger associations in nondiabetic individuals for some indices), most interaction p-values exceeded 0.05, indicating limited evidence of effect modification. In NHANES III, the associations remained directionally consistent across subgroups, further supporting the consistency of the findings across cohorts and subgroups.

Machine-learning modeling and model interpretability

We subjected 34 candidate variables to six feature-selection algorithms (Fig. 7A-F). The intersection yielded 12 core predictors (Fig. 7G): FAR, FHR, race, albumin, hemoglobin, serum iron, TIBC, ALP, creatinine, total cholesterol, WBC, and RBC. This set intentionally excluded variables with substantial overlap with the MAFLD definition (e.g., BMI, diabetes, hypertension) to avoid circularity and to assess the incremental value of ferritin-derived indices and routine laboratory markers.

We assessed the discriminative performance of multiple machine-learning algorithms across the training set, internal test set, and the NHANES III external validation set to select the optimal predictive model. In the training set, tree-based ensemble models (Random Forest, Ranger) showed perfect discrimination (AUC = 1.000), implying overfitting to the training data. The Gradient Boosting Machine (GBM) exhibited the highest discriminative ability with an AUC of 0.731 (95% CI: 0.711-0.750), surpassing SVM-Radial (0.725), NeuralNet (0.724), and Ranger (0.719). In the independent NHANES III external validation set, GBM and SVM-Radial showed similarly high discrimination, both outperforming Rangerand Random Forest. Given its comparatively consistent performance across phases, GBM was selected as the final algorithm (Fig. 8A).The comprehensive performance heatmap and radar plots further suggested that GBM achieved relatively balanced and prominent comprehensive performance in both internal and external phases (Fig. 8B, C).

To unveil the decision-making process of the final GBM model, we employed both global and local interpretability approaches. Global SHAP analysis (Fig. 9A) identified FAR, WBC, RBC, TIBC, and ALB as the top five predictors, collectively accounting for 82.1% of total importance. For individualized interpretation, LIME analyses of four representative cases (Fig. 9B) detailed how specific physiological alterations locally increased (red) or decreased (blue) predicted risk, with probabilities ranging from 0.706 to 0.913. Finally, we built a user-friendly web-based dynamic diagnostic calculator based on the final GBM model to facilitate individualized MAFLD risk assessment (Fig. 9C).

In summary, ferritin-derived composite indices, especially FAR and FHR, exhibit stable, nonlinear associations with MAFLD across two large U.S. cohorts. The associations increased most steeply at low-to-moderate index levels and flattened at higher levels. The GBM model based on 12 variables showed moderate discrimination in external validation and may warrant further evaluation as an auxiliary risk assessment tool.

Discussion

In two large U.S. cohorts spanning three decades, ferritin-derived composite indices – particularly FAR and FHR – were consistently and nonlinearly associated with MAFLD, with the associations increasing steeply at lower index levels and flattening thereafter. An externally validated GBM model incorporating these indices achieved moderate discrimination (AUC ≈ 0.699) and may offer a basis for an auxiliary risk stratification tool.

FAR and FHR emerged as the most consistently associated indices. FAR normalizes ferritin by albumin, which may reflect both hepatic synthetic capacity and nutritional status; FHR normalizes ferritin by hemoglobin, which may partly account for variation related to anemia. Both were consistently associated with MAFLD in development and external cohorts, ranked highly in GBM importance, and showed similar nonlinear patterns. By contrast, FALR and FLR – which additionally adjust for lymphocyte count – may incorporate variation in immune parameters associated with MAFLD, potentially attenuating their observed associations.

aThe GBM model’s external AUC of 0.699 indicates moderate discrimination but does not, by itself, establish incremental predictive value from non-definition variables – including ferritin-derived indices and routine laboratory markers – because we deliberately excluded BMI, diabetes, and hypertension to avoid overlap with the MAFLD definition. Thus, the model may complement, but should not replace, established diagnostic criteria. Its moderate performance may justify further evaluation for risk triage or screening, but does not establish clinical utility. SHAP and LIME help characterize feature contributions at both population and individual levels; the web-based calculator could facilitate rapid point-of-care risk estimation.

External validation in NHANES III confirmed directional reproducibility, but effect sizes were larger in the historical cohort. This likely reflects differences in steatosis measurement (ultrasound vs. CAP), secular trends in metabolic health, evolving laboratory assays, and application of development-derived cut-points that may not optimally fit the historical distribution. Quantitative comparisons across eras should be interpreted cautiously.

Limitations include the cross-sectional design, which precludes causal inference; differing MAFLD definitions between cohorts; potential residual confounding; and the model’s moderate AUC, suggesting limited discrimination. Whether additional biomarkers would improve model performance remains to be determined. Generalisability beyond U.S. populations and prospective validation of thresholds and model utility require further investigation.

Conclusions

In two large, nationally representative U.S. cohorts, ferritin-derived composite indices – especially FAR and FHR – are consistently associated with MAFLD status in nonlinear patterns. An externally validated, interpretable GBM model incorporating these indices achieves moderate discriminative performance (AUC ≈ 0.699) and may provide a basis for further development of a web-based risk assessment tool. These findings identify ferritin-derived indices as potential candidates for MAFLD risk stratification and provide a predictive framework that may complement existing diagnostic approaches. Future prospective studies should evaluate longitudinal predictive utility and determine whether these tools improve clinical decisions, patient outcomes, or resource utilization.

Data availability

All individual-level data analyzed in this study are publicly available as de-identified public-use files from the NCHS/CDC NHANES data portal for NHANES 2017–March 2020, August 2021–August 2023, and NHANES III. The corresponding documentation and access links are provided in references 32–34. No restricted-use data were used.

Disclosures

This research received no external funding.

NHANES and NHANES III were conducted under the applicable National Center for Health Statistics ethics review approvals, with documented participant consent. This analysis used publicly released, deidentified data. No new participants were recruited or contacted.

The authors declare no conflict of interest.


Supplementary material is available on the journal’s website.

References

  1. Rinella ME, Lazarus JV, Ratziu V, et al. A multisociety Delphi consensus statement on new fatty liver disease nomenclature. J Hepatol 2023; 79: 1542-1556.
  2. Chan KE, Koh TJL, Tang ASP, et al. Global prevalence and clinical characteristics of metabolic-associated fatty liver disease: a meta-analysis and systematic review of 10 739 607 individuals. J Clin Endocrinol Metab 2022; 107: 2691-2700.
  3. Younossi ZM, Golabi P, Paik JM, et al. The global epidemiology of nonalcoholic fatty liver disease (NAFLD) and nonalcoholic steatohepatitis (NASH): a systematic review. Hepatology (Baltimore, Md.) 2023; 77: 1335-1347.
  4. Mantovani A, Csermely A, Petracca G, et al. Non-alcoholic fatty liver disease and risk of fatal and non-fatal cardiovascular events: an updated systematic review and meta-analysis. Lancet Gastroenterol Hepatol 2021; 6: 903-913.
  5. Mantovani A, Petracca G, Beatrice G, et al. Non-alcoholic fatty liver disease and risk of incident chronic kidney disease: an updated meta-analysis. Gut 2022; 71: 156-162.
  6. Mantovani A, Petracca G, Beatrice G, et al. Non-alcoholic fatty liver disease and increased risk of incident extrahepatic cancers: a meta-analysis of observational cohort studies. Gut 2022; 71: 778-788.
  7. Rinella ME, Neuschwander-Tetri BA, Siddiqui MS, et al. AASLD Practice Guidance on the clinical assessment and management of nonalcoholic fatty liver disease. Hepatology (Baltimore, Md.) 2023; 77: 1797-1835.
  8. Li JH, Ma XY, Yi Y, et al. Association between serum ferritin levels and metabolic-associated fatty liver disease in adults: a cross-sectional study based on the NHANES. Curr Med Sci 2024; 44: 494-502.
  9. Yu G, Liu L, Qin T, et al. Associations of serum iron status with MAFLD and liver fibrosis in the USA: a nationwide cross-section study. Biol Trace Elem Res 2024; 202: 87-98.
  10. Yang N, Lu Y, Cao L, et al. The association between non-alcoholic fatty liver disease and serum ferritin levels in American adults. J Clin Lab Anal 2022; 36: e24225.
  11. Moris W, Verhaegh P, Jonkers D, et al. Hyperferritinemia in nonalcoholic fatty liver disease: iron accumulation or inflammation? Semin Liver Dis 2019; 39: 476-482.
  12. He H, Liao S, Zeng Y, et al. Causal relationships between metabolic-associated fatty liver disease and iron status: Two-sample Mendelian randomization. Liver Int 2022; 42: 2759-2768.
  13. Liu C, Chen Y, Zhang Z, et al. Iron status and NAFLD among European populations: A bidirectional two-sample mendelian randomization study. Nutrients 2022; 14.
  14. Sun K, Zhao JV, Nelson EAS, et al. Iron status and non-alcoholic fatty liver disease: A Mendelian randomization study. Nutrition (Burbank, Los Angeles County, Calif.) 2024; 118: 112295.
  15. Chen J, Li X, Ge C, et al. The multifaceted role of ferroptosis in liver disease. Cell Death Differ 2022; 29: 467-480.
  16. Srivastav SK, Mir IA, Bansal N, et al. Serum ferritin in metabolic syndrome-mechanisms and clinical applications. Pathophysiology 2022; 29: 319-325.
  17. Soeters PB, Wolfe RR, Shenkin A. Hypoalbuminemia: Pathogenesis and clinical significance. JPEN J Parenter Enteral Nutr 2019; 43: 181-193.
  18. Song Y, Guo W, Li Z, et al. Systemic immune-inflammation index is associated with hepatic steatosis: Evidence from NHANES 2015-2018. Front Immunol 2022; 13: 1058779.
  19. Liu K, Tang S, Liu C, et al. Systemic immune-inflammatory biomarkers (SII, NLR, PLR and LMR) linked to non-alcoholic fatty liver disease risk. Front Immunol 2024; 15: 1337241.
  20. Dai X, Zhou G, Xu L. Associations between red blood cell count and metabolic dysfunction-associated fatty liver disease (MAFLD). PLoS One 2022; 17: e0279274.
  21. National Center for Health Statistics. NHANES 2017-March 2020 pre-pandemic data: questionnaires, datasets, and related documentation. Centers for Disease Control and Prevention 2026. Available from: https://wwwn.cdc.gov/nchs/nhanes/continuousnhanes/default.aspx?Cycle=2017-2020.
  22. National Center for Health Statistics. NHANES August 2021-August 2023: Questionnaires, datasets, and related documentation. Centers for Disease Control and Prevention 2026. Available from: https://wwwn.cdc.gov/nchs/nhanes/continuousnhanes/default.aspx?Cycle=2021-2023.
  23. National Center for Health Statistics. NHANES III hepatic/gallbladder ultrasound and hepatic steatosis data documentation. Centers for Disease Control and Prevention 2026. Available from: https://wwwn.cdc.gov/nchs/data/nhanes3/34a/HGUHS.htm.
Copyright: © Clinical and Experimental Hepatology. This is an Open Access journal, all articles are distributed under the terms of the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). License (http://creativecommons.org/licenses/by-nc-sa/4.0/) enables reusers to distribute, remix, adapt, and build upon the material in any medium or format for noncommercial purposes only, and only so long as attribution is given to the creator. If you remix, adapt, or build upon the material, you must license the modified material under identical terms.
Share
without publication fees