Revised: March 11, 2026
Accepted: April 20, 2026
Published online: June 27, 2026
Processing time: 161 Days and 18.3 Hours
Noninvasive assessment of liver fibrosis is essential for the management of chronic hepatitis B (CHB), but existing tools have important limitations. Liver biopsy is invasive and unsuitable for repeated use, while commonly used serum indices show variable performance across populations and elastography is not universally available. We hypothesized that a laboratory-based machine-learning model using routinely available tests could provide consistent discrimination across independent clinical settings.
To develop and externally validate a laboratory-based machine-learning model for identifying significant liver fibrosis in CHB.
This multicenter cohort study included adults with CHB who underwent liver biopsy in two tertiary hospitals as a development cohort and one independent tertiary hospital as an external validation cohort. A population-based cohort from the National Health and Nutrition Examination Survey was used for exploratory evaluation with surrogate fibrosis labels. After rule-based harmonization, a stacking ensemble model was trained using routine laboratory variables and evaluated using discrimination, calibration, and decision curve analyses.
The model demonstrated good discrimination in the development cohort [area under the curve (AUC) = 0.853] and in the biopsy-confirmed external cohort (0.838), with acceptable calibration in both. Compared with aspartate aminotransferase to platelet ratio index and fibrosis-4 (FIB-4), the model showed comparable discrimination and similar net clinical benefit across clinically relevant thresholds. Within the FIB-4 indeterminate zone, the model achieved an AUC of 0.821 and reclassified 69.8% of patients into lower- or higher-risk categories with a gradient in observed fibrosis prevalence. In the population-based cohort with aspartate aminotransferase to platelet ratio index/FIB-4-based surrogate labeling, the model preserved risk ranking (AUC 0.817).
A routine laboratory-based machine-learning model provides stable discrimination for significant fibrosis in CHB across clinical settings, with explainability supported by feature attribution analysis.
Core Tip: In this study, a laboratory-based machine-learning model was developed and externally validated for noninvasive identification of significant fibrosis in patients with chronic hepatitis B using biopsy-confirmed multicenter cohorts. The model showed stable discrimination and acceptable calibration across independent hospitals, and retained risk-ranking ability in an exploratory population-based cohort with surrogate fibrosis labels. Because it relies only on routinely available laboratory tests, this model may serve as a practical complementary tool for fibrosis risk stratification, particularly in settings where elastography is unavailable or inconsistently applied.
- Citation: Wang TT, Chu YL, Lou YQ, Yang RY, Pu MM, Shan LJ, Huang L, Chen SS, Huang HJ. Routine laboratory model for identifying significant fibrosis in chronic hepatitis B. World J Hepatol 2026; 18(6): 119005
- URL: https://www.wjgnet.com/1948-5182/full/v18/i6/119005.htm
- DOI: https://dx.doi.org/10.4254/wjh.119005
Chronic hepatitis B (CHB) remains a major global health challenge. The World Health Organization estimates that 254 million people were living with chronic hepatitis B virus (HBV) infection in 2022, with approximately 1.2 million new infections annually[1]. Liver fibrosis is the principal determinant of CHB-related morbidity and mortality; progression from significant fibrosis to cirrhosis sharply increases the risks of decompensation, liver failure, and hepatocellular carcinoma[2-5]. Accurate identification of patients with significant fibrosis (≥ F2/S2) is therefore essential for antiviral treatment decisions and long-term risk stratification[6-9].
Despite guideline endorsement, existing fibrosis assessment tools have important limitations. Liver biopsy, although the reference standard, is invasive and resource-intensive, and its accuracy is compromised by sampling variability, making it impractical for repeated or large-scale use[7]. Although non-invasive tests such as aspartate aminotransferase (AST) to platelet ratio index (APRI), fibrosis-4 (FIB-4), and transient elastography are widely used, their diagnostic performance varies substantially across populations, and access to elastography remains limited in many HBV-endemic settings[10-14]. Imaging-based modalities such as transient elastography provide higher accuracy but are constrained by operator dependence, limited accessibility, and susceptibility to common clinical confounders[15-17]. No simple non-invasive approach consistently performs well across heterogeneous HBV populations.
Machine-learning methods have been explored for fibrosis assessment, but most published models still encounter methodological constraints and challenges in clinical translation. Many machine-learning fibrosis models are built in single-center datasets and show calibration drift and unstable thresholds when evaluated in external populations, an issue of practical concern for clinical deployment[18,19]. Although feature attribution methods such as SHapley Additive exPlanations (SHAP) can partially improve interpretability, many machine-learning models remain difficult to interpret in routine clinical practice. Moreover, feature-distribution shifts arising from demographic, biochemical, or disease-related heterogeneity further degrade performance across settings[18]. There is still a need for machine learning (ML) models that can generalize across cohorts, keep usable thresholds when moved between centers, and be interpreted by clinicians.
We developed a multi-center, biopsy-anchored ML model for non-invasive assessment of significant fibrosis in CHB. Using over 1700 biopsy-confirmed cases from multiple hospitals within the same geographic region, we constructed a harmonized development cohort and employed a stacking ensemble to integrate complementary classifiers. External validity was assessed in an independent biopsy-confirmed hospital cohort, and transportability was explored in National Health and Nutrition Examination Survey (NHANES) using surrogate fibrosis phenotyping. Model interpretability was ensured via surrogate modeling and SHAP-based analyses. Our aim was not to replace existing non-invasive test, but to develop a laboratory-based tool that remains stable across centers and could complement current fibrosis assessment strategies where elastography is unavailable.
This study adopted a multi-center, biopsy-anchored design comprising one development cohort, one biopsy-confirmed external validation cohort, and one population-based exploratory cohort (Figure 1). The development cohort was assembled from two tertiary hospitals in Zhejiang Province, Taizhou Hospital (2014-2019) and Zhejiang Provincial People’s Hospital (2011-2021). An independent biopsy-confirmed cohort from the First Affiliated Hospital of Zhejiang University (FAHZU) School of Medicine (2014-2019) served as external validation cohort. All hospital-based participants were adults (age ≥ 18 years) with CHB who underwent liver biopsy with histological fibrosis staging assessed using the Scheuer system. Liver biopsy specimens were obtained under ultrasound guidance using an 18-gauge automated biopsy needle. Specimens were fixed in formalin, embedded in paraffin, and stained with hematoxylin-eosin and reticulin. Biopsy adequacy was predefined as a specimen length ≥ 1.5 cm containing at least six intact portal tracts. All slides were independently evaluated by three experienced pathologists who were blinded to clinical data. In cases of discrepant assessments, re-evaluation was performed to reach consensus.
Patients were eligible for inclusion if they were adults (age ≥ 18 years) with CHB, defined by documented hepatitis B surface antigen positivity for at least six months prior to biopsy, and had available histological fibrosis staging according to the Scheuer system. Patients were excluded if biopsy information was incomplete or unavailable, or if essential laboratory variables required for baseline characterization and model construction were missing after preprocessing. Additional exclusion criteria included evidence of other causes of chronic liver disease, including hepatitis C virus infection, alcoholic liver disease, autoimmune liver disease, non-alcoholic fatty liver disease, prior liver transplantation, or hepatocellular carcinoma at baseline.
The detailed flow of patient inclusion and exclusion across cohorts is shown in Figure 1, and the availability and screening of candidate variables are summarized in Supplementary Table 1. Transient elastography measurements were not systematically available across participating centers during the study period and therefore were not included in model development or comparative analyses.
Patients with primary non-alcoholic fatty liver disease diagnosed by imaging, clinical history, or physician documentation as the dominant cause of liver injury were excluded. In patients with CHB, the presence of metabolic comorbidities (e.g., obesity, diabetes, dyslipidemia) was not considered an exclusion criterion unless fatty liver disease was clinically determined to be the dominant etiology of liver injury. This approach reflects real-world clinical overlap between CHB and metabolic dysfunction and allows evaluation of fibrosis risk within typical CHB populations rather than a metabolically “purified” subgroup.
Data from the NHANES 2011-2020 were used as a population-based exploratory cohort to assess model transportability under substantial epidemiological and laboratory shift. Adult participants (age ≥ 18 years) were included if they showed evidence of HBV exposure, defined by positivity for hepatitis B surface antigen or antibody to hepatitis B core antigen, capturing both chronic infection and resolved or occult exposure. This cohort was not intended to serve as a formal external validation cohort, as liver biopsy is unavailable in NHANES. NHANES data were used solely for exploratory external evaluation and were not involved in model training, feature selection, or parameter estimation.
Because liver biopsy is not available in NHANES, fibrosis status could not be directly observed and was therefore approximated using a conservative, pre-specified surrogate phenotyping algorithm based on APRI and FIB-4[20,21]. Specifically, APRI < 0.6 was used to define non-significant fibrosis, APRI ≥ 1.0 to define significant fibrosis, and individuals in the intermediate APRI range (0.6-1.0) were further classified using FIB-4 ≥ 2.67; remaining indeterminate cases were excluded to retain high-confidence surrogate labels[22]. This cohort was used solely for external evaluation of model transportability, and NHANES data were not used for model training, feature selection, or parameter estimation. Across all cohorts, demographic variables, laboratory measurements, and HBV-related biomarkers were harmonized through a unified preprocessing pipeline.
Baseline demographic characteristics and routinely reported laboratory measurements were extracted from each cohort to describe the study populations (Table 1). These variables included age, sex, body mass index, liver aminotransferases [alanine aminotransferase (ALT) and AST], cholestatic enzymes [gamma-glutamyl transferase (GGT) and alkaline phosphatase (ALP)], bilirubin, albumin and globulin fractions, platelet count, lipid profiles, and selected complete blood count parameters.
| Variable1 | Development | External (First Affiliated Hospital of Zhejiang University) | National Health and Nutrition Examination Survey (population-based) | |
| Clinical features | Age (years) | 41.0 (16.0) | 38.0 (17.0) | 48.5 (18.8) |
| Sex (male) | 765 (65.6) | 367 (59.1) | 1284 (54.1) | |
| Body mass index (kg/m2) | 23.1 (4.4) | 22.2 (4.2) | 24.7 (5.6) | |
| Height (cm) | 167.0 (11.5) | 168.0 (12.0) | 165.3 (14.2) | |
| Weight (kg) | 64.0 (14.8) | 61.0 (16.0) | 73.2 (23.4) | |
| Hepatitis B virus virology | Hepatitis B surface antigen, log10 (IU/mL) | 2.4 (1.4) | 3.6 (1.7) | |
| Liver function tests | Alanine aminotransferase (U/L) | 36.0 (35.0) | 43.0 (58.0) | 57.7 (55.3) |
| Aspartate aminotransferase (U/L) | 32.0 (22.0) | 31.0 (29.0) | 35.0 (23.5) | |
| Gamma-glutamyl transferase (U/L) | 27.0 (28.0) | 27.0 (38.0) | 21.0 (18.0) | |
| Alkaline phosphatase (U/L) | 87.5 (36.0) | 72.0 (32.0) | 67.0 (28.0) | |
| Albumin (g/L) | 43.1 (5.6) | 44.7 (7.0) | 38.1 (3.5) | |
| Globulin (g/L) | 28.1 (5.6) | 26.6 (5.6) | 30.0 (10.0) | |
| White blood cell count (109/L) | 5.4 (2.1) | 5.4 (2.1) | 5.1 (6.5) | |
| Red blood cell count (1012/L) | 4.7 (0.8) | 4.7 (0.8) | 4.6 (0.7) | |
| Blood routine characteristics | Platelet count (109/L) | 176.0 (78.0) | 190.0 (71.0) | 205.5 (80.5) |
| Mean corpuscular volume (fL) | 91.0 (5.8) | 90.2 (5.5) | 89.7 (7.0) | |
| Total cholesterol (mmol/L) | 4.3 (1.2) | 4.1 (1.2) | 3.9 (0.7) | |
| Metabolic indicators | High density lipoprotein-cholesterol (mmol/L) | 1.2 (0.4) | 1.2 (0.5) | 1.1 (0.3) |
| Low-density lipoprotein-cholesterol (mmol/L) | 2.5 (0.9) | 2.3 (1.0) | 2.4 (0.6) |
In addition to variables used for baseline description, a broader set of candidate predictors was assembled for model development. To ensure cross-cohort comparability, all candidate variables were processed using a pre-specified, rule-based harmonization pipeline applied before any modeling. Harmonization consisted of three steps: (1) Semantic harmonization. Variables representing the same clinical construct were mapped to a unified standardized name and definition (e.g., aligning “platelet/PLT” across centers; consistent coding for sex and HBV serology); (2) Unit harmonization. Laboratory units were converted to consistent units across cohorts where necessary, and unit-inconsistent entries were treated as invalid values; and (3) Definition harmonization and derivation. When an equivalent construct was not directly measured in a cohort, it was derived only when a clinically defensible definition existed and required components were available (e.g., globulin = total protein - albumin in NHANES). Variables without a defensible mapping were not considered for cross-cohort modeling.
Pre-specified outlier handling. Clinically implausible values were identified using conservative, pre-specified plausibility rules defined a priori for each variable (Supplementary Table 2). Values outside these physiologically plausible ranges were set to missing prior to imputation. These rules were applied independently of outcome information and were intended to remove measurement or data-entry errors rather than restrict disease severity. Outlier handling was performed without using outcome information and before any model fitting.
To minimize selection bias, we did not perform a complete-case analysis. Missingness-related exclusion was applied only in the presence of structural missingness, defined a priori as the absence of ≥ 3 of the five core predictors required for minimal model input (age, platelet count, albumin, AST, and ALT) at baseline, which would preclude reliable reconstruction of clinically derived features and make imputation unstable. For all remaining missing values, imputation was performed using fold-specific median imputation within each training fold, and continuous variables were standardized using z-scores. All preprocessing steps (outlier handling → imputation → scaling) were implemented in a leakage-free manner: Preprocessing parameters were estimated from the corresponding training split and then applied unchanged to the held-out split. To assess robustness to the imputation strategy, we conducted a prespecified sensitivity analysis using alternative imputation approaches, including k-nearest neighbors and multiple imputation by chained equations, within the same fold-specific, leakage-free framework. Model development and evaluation were repeated without altering the feature set, algorithms, or operating point. Results are summarized in Supplementary Table 3.
Model development overview: Model development followed a pre-specified, leakage-free workflow designed to maximize cross-cohort transportability and clinical deploy ability. Candidate predictors were defined a priori based on routinely available clinical information, including demographic characteristics, liver biochemistry, hematologic indices, lipid markers, and a small number of clinically motivated derived ratios, such as the AST/ALT ratio and albumin/globulin (A/G) ratio. All candidate variables were restricted to measurements that are commonly reported in routine practice and could be harmonized across centers. Variable availability across cohorts, derivation rules, and final inclusion are detailed in Supplementary Table 2.
Pre-modeling feasibility screening: Prior to model fitting, candidate variables underwent a feasibility screening process to ensure that retained features could be consistently applied across heterogeneous clinical datasets. This screening was conducted independently of outcome information and was based on pre-defined, objective criteria: (1) Availability and mappability: Variables not measured, or not meaningfully derivable using standard clinical definitions, across all hospital-based cohorts were excluded; (2) Harmonizability and comparability: Variables with incompatible measurement definitions, assay characteristics, or clinical interpretations across cohorts, precluding reliable harmonization, were removed; and (3) Missingness assessment: Variable-level missingness was assessed within hospital-based cohorts. A conservative missingness threshold (> 20%) was applied as a secondary exclusion criterion; however, for most routinely reported laboratory variables, missingness was low and was not the primary driver of exclusion.
After harmonization and feasibility screening, model training proceeded using a stacking ensemble framework. SHAP-based importance analysis was used primarily to aid interpretation and to guide parsimony evaluation rather than as a sole feature-selection method. This step was used to improve parsimony and interpretability while preserving predictive performance.
Based on this process, the final model incorporated 13 routinely available laboratory features. All retained predictors were either directly measured or constructed using pre-defined clinical formulas and were available across all evaluated datasets, including NHANES. Variables removed included lactate dehydrogenase, alpha-fetoprotein, alpha-L-fucosidase, HBV DNA, and several metabolic and electrolyte markers, primarily due to incomplete availability or lack of harmonizable definitions across cohorts. Final predictors included age, platelet count, albumin, globulin, A/G, AST, AST/ALT ratio, ALP, GGT, white blood cell count, red blood cell count (RBC), mean corpuscular volume (MCV), and low-density lipoprotein (LDL)-cholesterol. Multicollinearity among the final predictors was assessed prior to model fitting. Pairwise dependencies were examined using Spearman correlation coefficients, and variance inflation factors (VIFs) were calculated to quantify multivariable linear dependency. As expected, a strong inverse correlation was observed between globulin and the A/G due to their mathematical relationship. However, all VIF values remained below conventional thresholds, indicating no evidence of problematic multicollinearity. Detailed correlation results and VIF values are provided in Supplementary Figure 1 and Supplementary Table 4.
To justify algorithm selection, we considered the characteristics of routine laboratory data and the intended deployment setting. Routine biomarkers exhibit non-linear effects, interactions, and heterogeneous scaling across sites, while the feature dimension is modest and the outcome is binary. We therefore included complementary model families with different inductive biases: (1) Logistic regression as a transparent linear baseline that is well-calibrated and robust under limited feature sets; (2) Tree-based ensembles (random forest and gradient boosting) to capture non-linearities and high-order interactions without requiring explicit specification; (3) Support vector machines as margin-based classifiers that can perform well in moderate-dimensional settings and provide a different decision boundary geometry; and (4) CatBoost as a modern gradient-boosting implementation with strong performance and stability in tabular clinical data.
We used an ensemble strategy because no single algorithm is consistently optimal across heterogeneous cohorts, where differences in laboratory platforms, patient mix, and disease spectrum can alter the relative performance of model classes. A stacking ensemble combines models that make partially uncorrelated errors, which can reduce variance and mitigate model-specific biases, thereby improving robustness under dataset shift. Importantly, all base learners were trained under the same leakage-free cross-validation procedure, and the ensemble was finalized within the development cohort before being evaluated unchanged in external cohorts.
The final model was implemented as a two-layer stacking ensemble. In the first layer, six base learners, logistic regression (L2-regularized), random forest, gradient boosting, support vector machine (RBF kernel), extra trees, and CatBoost, were trained under stratified five-fold cross-validation using the same leakage-free preprocessing pipeline. Within each fold, each base learner was fitted on the training split and generated out-of-fold (OOF) predicted probabilities for the held-out split. These OOF probabilities from all base learners were concatenated to form a level-1 meta-feature matrix.
The second-layer meta-learner was a logistic regression model (L2-regularized; solver = LBFGS) trained only on the level-1 OOF meta-features, thereby preventing information leakage and providing an unbiased internal estimate of ensemble performance. After cross-validation, base learners were refit on the full development cohort using the finalized preprocessing procedure, and the meta-learner was applied to the corresponding stacked probability outputs to produce the final continuous predicted probability of significant fibrosis. Model performance was then evaluated using discrimination, calibration, and decision curve analyses. Full hyperparameter settings and training details are provided in Supplementary Table 5.
Multiple complementary base learners, including logistic regression, random forest, gradient boosting models, support vector machines, and CatBoost, were trained using stratified five-fold cross-validation within the development cohort. OOF predictions from individual base learners were combined using a meta-learner to generate the final ensemble probability for significant fibrosis.
All preprocessing steps, including missing-value imputation and standardization of continuous variables, were performed strictly within each training fold and applied to the corresponding validation fold to prevent information leakage. External validation cohorts were processed using the finalized harmonization rules, preprocessing parameters, and trained model, without incorporating any external data into feature selection or parameter estimation. This study was conducted and reported in accordance with the TRIPOD reporting guideline for prediction model development and validation.
Across the three cohorts, participants demonstrated demographic and biochemical heterogeneity, reflecting differences in study setting and cohort composition (Table 1). The development cohort (n = 1166) had a median age of 41.0 years, similar to external validation cohort (FAHZU; median 38.0 years) but younger than population-based exploratory cohort (NHANES; median 48.5 years). Body mass index was highest in NHANES (24.7 kg/m2), consistent with its population-based design.
Biochemical characteristics varied across cohorts. Liver enzyme levels, particularly ALT and AST, tended to be higher in NHANES compared with the hospital-based cohorts, whereas GGT levels were comparable or lower. Serum albumin concentrations were highest in FAHZU and lowest in NHANES, while globulin levels were modestly higher in NHANES. Hematological indices were broadly comparable across cohorts. Platelet counts were slightly higher in NHANES, whereas white and RBCs showed similar distributions. Lipid profiles exhibited cross-cohort differences, with lower total cholesterol, high density lipoprotein-cholesterol, and LDL-cholesterol levels observed in NHANES relative to the development and FAHZU cohorts.
Substantial cross-cohort heterogeneity was observed in the statistical distributions of key laboratory variables (Supplementary Figures 2 and 3). Kolmogorov-Smirnov tests (Supplementary Figure 2A) demonstrated significant distributional differences between the development cohort and the biopsy-confirmed external cohort (FAHZU) as well as the NHANES exploratory cohort for most features, including liver enzymes (ALT, AST, GGT), platelet count, globulin fractions, and lipid markers, indicating non-trivial dataset shift. Only a small subset of variables, such as MCV, showed partial overlap across cohorts.
In contrast to distributional differences, all final model predictors were available across all cohorts after harmonization, and no clinically relevant systematic missingness was observed (Supplementary Figure 2B). However, NHANES exhibited measurement-platform differences typical of population-based surveys, for example, wider value ranges for metabolic markers and differing central tendencies for globulin-related variables, reflecting its distinct sampling frame and laboratory protocols, and underscoring that NHANES serves as a heterogeneous, population-based setting for exploratory model evaluation.
Visual comparisons of feature distributions (Supplementary Figure 3) further highlighted this structural heterogeneity. NHANES participants were older, had lower globulin and GGT concentrations, and showed broader variability in platelet counts and lipid profiles, while FAHZU displayed higher dispersion in liver enzyme values than the development cohort.
To further evaluate clinical utility in cases where conventional serum indices are inconclusive, we examined model performance among patients within the FIB-4 indeterminate zone (n = 129) (Table 2). Within this subgroup, the model preserved ranking discrimination, achieving an area under the curve (AUC) of 0.821 [95% confidence interval (CI): 0.737-0.895, Figure 2].
| Group | n | S ≥ 21 | S ≥ 2 (%) |
| Low risk | 54 | 7 | 13.0 |
| Uncertain | 39 | 14 | 35.9 |
| High risk | 36 | 27 | 75.0 |
Risk stratification within the indeterminate zone demonstrated a clear gradient in the observed prevalence of significant fibrosis (S ≥ 2). Among individuals assigned to the lower-risk stratum, 7 of 54 (13.0%) had significant fibrosis. In the intermediate stratum, 14 of 39 (35.9%) had significant fibrosis, whereas in the higher-risk stratum, 27 of 36 (75.0%) were confirmed to have significant fibrosis. Overall, 90 of 129 (69.8%) indeterminate cases were redistributed into lower- or higher-risk categories, leaving 39 (30.2%) in an intermediate range.
At the same prespecified probability setting applied in the primary analysis, sensitivity and specificity within the indeterminate subgroup were 0.854 and 0.617, respectively. The negative predictive value was 0.877, with an F1-score of 0.683 and overall accuracy of 0.705. These findings indicate that even among patients for whom FIB-4 alone does not provide a definitive classification, the laboratory-based model preserves coherent risk ordering and meaningfully separates individuals with differing probabilities of significant fibrosis.
The ensemble model demonstrated strong discriminative performance in the development cohort, achieving an AUC of 0.853 (95%CI: 0.830-0.872) with sensitivity 0.809 (95%CI: 0.771-0.845), specificity 0.708 (95%CI: 0.674-0.739), positive predictive value 0.621, negative predictive value 0.862, and F1-score 0.703 (95%CI: 0.671-0.734) (Figure 3; Table 3).
| Cohort1 | AUC (95%CI) | Sensitivity (95%CI) | Specificity (95%CI) | PPV | NPV | F1 (95%CI) |
| Development | 0.8528 (0.8301-0.8722) | 0.8088 (0.7711-0.8449) | 0.7077 (0.6735-0.7394) | 0.6212 | 0.8619 | 0.7027 (0.6708-0.7337) |
| External (First Affiliated Hospital of Zhejiang University) | 0.8379 (0.8000-0.8721) | 0.7290 (0.6506-0.7949) | 0.7361 (0.6966-0.7778) | 0.4788 | 0.8909 | 0.5780 (0.5151-0.6383) |
| National Health and Nutrition Examination Survey (surrogate-labeled) | 0.817 (0.800-0.833) |
When evaluated in the biopsy-confirmed external cohort (FAHZU), the model maintained comparable discrimination (AUC 0.838, 95%CI: 0.800-0.872), with sensitivity 0.729 (95%CI: 0.651-0.795), specificity 0.736 (95%CI: 0.697-0.778), positive predictive value 0.479, negative predictive value 0.891, and F1-score 0.578 (95%CI: 0.515-0.638), supporting its transportability across independent hospital settings with differing laboratory platforms and patient characteristics (Figure 3; Table 3).
In the population-based NHANES cohort, where fibrosis status was defined using APRI/FIB-4-based surrogate criteria rather than histology, the model maintained discrimination (AUC 0.817) despite substantial epidemiological and biochemical heterogeneity and surrogate outcome definition. (Figure 3; Table 3). Because surrogate labeling does not support a single clinically transportable operating point, AUC was considered the primary metric in NHANES. Exploratory classification performance under two probability settings (original model output and a simple probability mapping) is provided in Supplementary Table 6 and should be interpreted cautiously as specific to the surrogate-defined phenotype. Performance was consistent across imputation methods (median, k-nearest neighbors, and multiple imputation by chained equations), with only minimal differences in discrimination, F1-score, and Brier score in both the development and external cohorts (Supplementary Table 3).
Calibration was evaluated in the development cohort and the biopsy-confirmed external cohort (FAHZU) using calibration curves and summary statistics (Figure 4). In the development cohort, predicted probabilities showed close agreement with observed event rates, with a calibration slope close to 1 and minimal intercept deviation, indicating limited over- or under-estimation of absolute risk. In FAHZU, the calibration curve preserved monotonicity (higher predicted risk corresponded to higher observed fibrosis prevalence), while a modest shift in calibration-in-the-large was observed, consistent with differences in baseline fibrosis prevalence and case-mix between centers. Overall, these findings support the model’s use for risk stratification in independent hospital settings. If site-specific absolute risk estimation is required for implementation, a simple recalibration (e.g., intercept update or logistic recalibration) using local data could be applied without changing discrimination. Overall, calibration curves demonstrated preserved monotonicity and supported the model’s use for clinical risk stratification.
Subgroup analyses in external validation cohort (FAHZU) demonstrated stable discriminatory performance across key clinical strata (Supplementary Figure 4). The AUCs across subgroups defined by age, sex, body mass index, ALT level (above vs below median), and platelet count remained close to the overall cohort AUC (0.838), with overlapping 95%CIs and no subgroup showing a marked decline in performance.
Notably, discrimination was preserved in patients with higher ALT levels and in those with lower platelet counts, supporting the robustness of the model in clinically relevant higher-risk subgroups within a real-world hospital setting. Precision-recall curves (Supplementary Figure 5) showed consistent performance across cohorts, with average precision values concordant with receiver operating characteristic-based discrimination, supporting the robustness of the model under class-imbalanced conditions.
Decision curve analysis (DCA) was used to estimate net benefit across threshold probabilities (Figure 5). We prespecified 0.20-0.50 as the primary interpretive range because, in routine practice, clinicians typically consider additional fibrosis work-up (e.g., elastography when available, closer follow-up, or biopsy in selected cases) within a moderate risk window rather than at extreme thresholds. In the development cohort, the ensemble model provided net benefit over treat-all and treat-none strategies across most of the prespecified range and was comparable to APRI and FIB-4. In the external biopsy-confirmed cohort (FAHZU), the ensemble curve largely overlapped with APRI and FIB-4 within 0.20-0.50, indicating similar clinical utility rather than a large incremental gain. Importantly, the ensemble model was not inferior to treat-all or treat-none within the prespecified range, supporting its role as a pragmatic laboratory-based triage tool where elastography is unavailable or inconsistently applied.
In the development cohort, the ensemble model demonstrated a higher net benefit than treat-all and treat-none strategies across clinically relevant threshold probabilities, and showed comparable or slightly greater net benefit compared with APRI and FIB-4. In external validation cohort (FAHZU), within the 0.20-0.50 threshold range, the decision curves of the ensemble model largely overlapped with those of APRI and FIB-4. Although the absolute magnitude of net benefit was similar, the ensemble model did not show inferior performance at any clinically relevant threshold and did not yield net benefit lower than treat-all or treat-none strategies. In some threshold ranges, the ensemble model yielded similar net benefit, suggesting limited but consistent net benefit in selected threshold ranges.
For NHANES, where fibrosis labels were derived using surrogate APRI/FIB-4 criteria (Supplementary Figure 6), the DCA reflects proxy performance rather than biopsy-validated outcomes. Even under surrogate labeling, the model performed no worse than simple treat-all or treat-none strategies across most thresholds, indicating potential exploratory utility in population-based samples.
To facilitate clinical interpretation, model explainability was assessed using SHAP analysis based on a surrogate model (Figure 6; Supplementary Figure 7). Global feature importance indicated that platelet count was the most influential predictor, followed by the A/G ratio, AST/ALT ratio, and age. These variables are well-established markers of portal hypertension, synthetic liver function, hepatocellular injury, and disease duration, respectively.
SHAP summary plots demonstrated clinically intuitive directions of effect. Lower platelet counts and lower A/G ratios were consistently associated with higher predicted risk of significant fibrosis, while higher AST/ALT ratios and older age shifted predictions toward fibrosis. Additional contributors included ALP, GGT, AST, globulin, albumin, RBC, white blood cell count, MCV, and LDL-cholesterol, each showing smaller but consistent effects.
Feature dependence analyses further supported these patterns, showing non-linear relationships consistent with known fibrosis biology. For example, predicted risk increased markedly at lower platelet counts and lower A/G ratios, while age and AST/ALT ratio exhibited gradual monotonic effects. Interaction analyses suggested that the impact of individual markers varied across clinical contexts, such as platelet level or enzyme ratio, without contradicting established clinical understanding. Overall, these results indicate that the model’s predictions are driven primarily by routine laboratory variables with clear clinical relevance and behave in directions consistent with known mechanisms of fibrosis progression, supporting the interpretability and credibility of the model in clinical settings.
Performance of individual base learners and the final stacking ensemble is summarized in Supplementary Table 7. Across cohorts, single learners showed variable discrimination, whereas the stacking ensemble provided a balanced and stable performance profile, supporting its selection as the final model. Tree-based models such as random forest and extra trees achieved high AUCs in the development cohort, but their discrimination decreased to varying degrees when applied to external datasets, particularly in the population-based NHANES cohort. Gradient boosting showed pronounced performance degradation under cross-cohort shift.
In contrast, the stacking ensemble achieved consistently strong discrimination across all cohorts, without being the top-performing model in any single dataset. This stability suggests that model aggregation mitigated algorithm-specific biases and sensitivity to cohort heterogeneity, resulting in improved transportability. Based on its balanced performance across development and external validation cohorts, the stacking ensemble was selected as the final model.
In this multi-center study, we developed and externally validated a laboratory-based machine-learning model for non-invasive identification of significant fibrosis in patients with CHB. Using biopsy-confirmed cohorts from three tertiary hospitals, the model demonstrated stable and reproducible discrimination across independent clinical settings. When further evaluated in a population-based cohort derived from NHANES, the model maintained ranking ability under surrogate-defined fibrosis labels. Importantly, this analysis should be interpreted as an assessment of score transportability under substantial covariate shift rather than an external validation of diagnostic accuracy, because fibrosis status in NHANES was not biopsy-confirmed and was approximated using APRI/FIB-4–based surrogate phenotyping. These findings suggest that the proposed model captures routinely measured fibrosis-related laboratory signals that remain informative across different clinical contexts.
Accurate assessment of liver fibrosis is central to the management of CHB, guiding antiviral treatment decisions and long-term risk stratification. Although liver biopsy remains the reference standard, its invasiveness and limited feasibility restrict routine use. Non-invasive tests such as APRI and FIB-4 are widely adopted but show variable performance across populations and clinical settings, while access to elastography remains inconsistent in many regions. In this context, our results indicate that a model based solely on routinely available laboratory tests can achieve discrimination comparable to, and in some settings exceeding, conventional serum indices, while maintaining stability across centers. In practice, such a laboratory-based model may function as a triage tool for initial fibrosis risk stratification, helping identify patients who may benefit from further diagnostic evaluation. Importantly, the model retained discriminative capacity within the FIB-4 indeterminate zone, where conventional serum indices fail to provide definitive classification. In this subgroup, the model achieved an AUC of 0.821 and reclassified approximately 70% of patients into lower- or higher-risk categories with a clear gradient in observed fibrosis prevalence. This suggests potential utility as a second-line tool in cases where FIB-4 yields inconclusive results.
Beyond discrimination, calibration is essential for clinical applicability. In the biopsy-confirmed cohorts, the model showed acceptable calibration, with close agreement between predicted and observed risks. In the external hospital cohort, an intercept shift was observed, indicating modest miscalibration in overall risk level (calibration-in-the-large). A negative calibration intercept suggests that the model slightly overestimated absolute risk in this cohort, which is expected when transporting a model to a population with different baseline fibrosis prevalence and case-mix characteristics. Importantly, the calibration slope and curve monotonicity were preserved across clinically relevant probability ranges, supporting the model’s use for risk stratification. If site-specific absolute risk estimates are required, a simple intercept update or logistic recalibration could be applied using local data without altering discrimination.
Subgroup analyses further supported the robustness of the model in real-world clinical practice. Discriminatory performance remained stable across subgroups defined by age, sex, body mass index, ALT level, and platelet count, with overlapping CIs and no clinically meaningful loss of performance. Notably, the model retained discrimination in patients with elevated ALT levels and in those with lower platelet counts, groups often associated with more advanced disease, suggesting applicability across a broad spectrum of disease severity.
DCA is most informative within clinically plausible threshold probabilities; extreme thresholds are rarely used in practice and can yield unstable net-benefit estimates; therefore our interpretation focuses on the moderate range where clinical decisions would realistically be made. While the magnitude of improvement over existing serum scores was modest, the model consistently outperformed treat-all and treat-none strategies. These findings support a complementary role for the model, particularly in settings where elastography is unavailable or inconsistently applied. In NHANES, where fibrosis status was defined using surrogate APRI/FIB-4 criteria, decision curve results should be interpreted cautiously and viewed as exploratory rather than definitive. Therefore, the NHANES findings should be viewed as evidence that the laboratory-based score preserves risk ordering across a markedly different sampling frame and laboratory platform, but they do not establish histological diagnostic accuracy or support clinical decision-making in community populations without further biopsy-anchored validation.
Interpretability analyses indicated that the model’s predictions were primarily driven by established clinical markers of fibrosis, including platelet count, A/G ratio, AST/ALT ratio, and age. Platelet count reflects portal hypertension and splenic sequestration associated with fibrosis progression, whereas a reduced A/G ratio reflects impaired hepatic synthetic function and chronic inflammatory activity. Their effects followed clinically intuitive directions, supporting the biological plausibility of the model and enhancing its transparency for clinical use. In addition, model performance remained stable across alternative imputation strategies, suggesting that predictive performance was not driven by a specific missing-data handling approach.
Several limitations merit consideration. First, the NHANES analysis is exploratory because fibrosis status was not biopsy-confirmed. Labels were derived from APRI/FIB-4-based surrogate criteria, which introduces unavoidable circularity: Several core predictors used by our model (e.g., age, AST, ALT, and platelet count) also enter the surrogate definitions. As a result, performance in NHANES primarily reflects preservation of risk ranking relative to the surrogate phenotype and should not be interpreted as evidence of diagnostic accuracy for histological fibrosis. Second, residual inter-cohort differences may persist despite harmonization. Third, although the model relies on routine laboratory variables, inter-laboratory variability may affect absolute risk estimates. Fourth, transient elastography data were not systematically available across participating centers during the retrospective study period. As a result, we were unable to perform a direct head-to-head comparison between the proposed laboratory-based model and liver stiffness measurement. Therefore, the incremental value of the model relative to elastography cannot be established in the current study. The intended positioning of this model is not to compete with elastography in tertiary centers where it is readily accessible, but rather to serve as a complementary or alternative tool in settings where elastography is unavailable, inconsistently applied, or not routinely recorded. Prospective studies incorporating simultaneous elastography and laboratory assessment are needed to clarify their relative and combined utility. Furthermore, residual overlap between CHB and metabolic dysfunction-associated steatotic liver disease cannot be entirely excluded, and metabolic factors may contribute to fibrosis progression in some patients. However, this reflects real-world clinical populations rather than an artificial exclusion of common comorbidities. Finally, the retrospective design underscores the need for prospective validation before clinical implementation.
Future work should include validation in broader clinical environments, especially primary-care and outpatient hepatology settings where access to biopsy or elastography is constrained[23-25]. Evaluation in lower-resource or mixed-care systems will further clarify real-world transportability. Incorporating multimodal data, such as hepatitis b core-related antigen, Mac-2 binding protein glycosylation isomer[26-31], ultrasound elastography, or imaging-derived features, may improve biological specificity and mitigate distributional shift[32]. Integration of the score into electronic health record systems, with automated output and SHAP-based explanations, could facilitate real-time clinical decision support. Head-to-head comparisons with recently published ML fibrosis models and assessment of downstream clinical impact will also be essential.
In this multi-center, biopsy-anchored study, we developed and externally validated a laboratory-based machine-learning score for identifying significant fibrosis in patients with CHB. The model demonstrated stable discrimination and acceptable calibration across independent cohorts while relying exclusively on routinely available laboratory variables. As a continuous risk stratification tool, it may complement existing non-invasive assessments and help identify patients who warrant further evaluation, particularly in clinical settings where elastography is unavailable or inconsistently applied. Prospective validation across broader clinical environments will be required to determine its role in routine clinical practice.
| 1. | World Health Organization. Hepatitis B. Jul 23, 2025. [cited 21 November 2025]. Available from: https://www.who.int/news-room/fact-sheets/detail/hepatitis-b. |
| 2. | Terrault NA, Lok ASF, McMahon BJ, Chang KM, Hwang JP, Jonas MM, Brown RS Jr, Bzowej NH, Wong JB. Update on prevention, diagnosis, and treatment of chronic hepatitis B: AASLD 2018 hepatitis B guidance. Hepatology. 2018;67:1560-1599. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 3558] [Cited by in RCA: 3220] [Article Influence: 402.5] [Reference Citation Analysis (8)] |
| 3. | Kim HL, Kim GA, Park JA, Kang HR, Lee EK, Lim YS. Cost-effectiveness of antiviral treatment in adult patients with immune-tolerant phase chronic hepatitis B. Gut. 2021;70:2172-2182. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 18] [Cited by in RCA: 39] [Article Influence: 7.8] [Reference Citation Analysis (0)] |
| 4. | Yan R, Sun M, Yang H, Du S, Sun L, Mao Y. 2024 latest report on hepatitis B virus epidemiology in China: current status, changing trajectory, and challenges. Hepatobiliary Surg Nutr. 2025;14:66-77. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 32] [Cited by in RCA: 42] [Article Influence: 42.0] [Reference Citation Analysis (0)] |
| 5. | Ou TY, Huy LD, Mayne J, Shih CL, Mai Xuan H, Thi Hong Nguyen N, Nguyen Hoai L, Thi My Bui L, Chang YM, Abdi AA, Hsu SC, Lin HJ, Huang CC. Global mortality of chronic liver diseases attributable to Hepatitis B virus and Hepatitis C virus infections from 1990 to 2019 and projections to 2030. J Infect Public Health. 2024;17:102443. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 37] [Cited by in RCA: 31] [Article Influence: 15.5] [Reference Citation Analysis (1)] |
| 6. | Shen Y, Wu SD, Chen Y, Li XY, Zhu Q, Nakayama K, Zhang WQ, Weng CZ, Zhang J, Wang HK, Wu J, Jiang W. Alterations in gut microbiome and metabolomics in chronic hepatitis B infection-associated liver disease and their impact on peripheral immune response. Gut Microbes. 2023;15:2155018. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 84] [Cited by in RCA: 68] [Article Influence: 22.7] [Reference Citation Analysis (0)] |
| 7. | Zheng P, Dou Y, Wang Q. Immune response and treatment targets of chronic hepatitis B virus infection: innate and adaptive immunity. Front Cell Infect Microbiol. 2023;13:1206720. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 31] [Cited by in RCA: 35] [Article Influence: 11.7] [Reference Citation Analysis (1)] |
| 8. | Ginès P, Castera L, Lammert F, Graupera I, Serra-Burriel M, Allen AM, Wong VW, Hartmann P, Thiele M, Caballeria L, de Knegt RJ, Grgurevic I, Augustin S, Tsochatzis EA, Schattenberg JM, Guha IN, Martini A, Morillas RM, Garcia-Retortillo M, de Koning HJ, Fabrellas N, Pich J, Ma AT, Diaz MA, Roulot D, Newsome PN, Manns M, Kamath PS, Krag A; LiverScreen Consortium Investigators. Population screening for liver fibrosis: Toward early diagnosis and intervention for chronic liver diseases. Hepatology. 2022;75:219-228. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 249] [Cited by in RCA: 207] [Article Influence: 51.8] [Reference Citation Analysis (0)] |
| 9. | Sun Y, Chen W, Chen S, Wu X, Zhang X, Zhang L, Zhao H, Xu M, Chen Y, Piao H, Li P, Li L, Jiang W, Li X, Xing H, Liu X, Zhang Y, Wang B, Zhou J, Meng T, Zhao X, Shao C, Kong Y, Zhao X, Ou X, Liu C, Jia J, You H. Regression of Liver Fibrosis in Patients on Hepatitis B Therapy Is Associated With Decreased Liver-Related Events. Clin Gastroenterol Hepatol. 2024;22:591-601.e3. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 13] [Cited by in RCA: 17] [Article Influence: 8.5] [Reference Citation Analysis (3)] |
| 10. | Yan P, Yu X, Chen Z, Lan L, Kang J, Zhao B, Liu D. Assessing the consistency of FIB-4, APRI, and GPR in evaluating significant liver fibrosis and cirrhosis in COVID-19 patients with concurrent liver diseases. BMC Gastroenterol. 2025;25:191. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in RCA: 4] [Reference Citation Analysis (0)] |
| 11. | Moosavy SH, Eftekhar E, Davoodian P, Nejatizadeh A, Shadman M, Zare S, Nazarnezhad MA. AST/ALT ratio, APRI, and FIB-4 compared to FibroScan for the assessment of liver fibrosis in patients with chronic hepatitis B in Bandar Abbas, Hormozgan, Iran. BMC Gastroenterol. 2023;23:145. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 15] [Reference Citation Analysis (0)] |
| 12. | Danis N, Gunsar F, Yilmaz F, Nart D, Turan I, Karasu Z, Ersoz G, Akarca US, Ozutemiz O. Performance of non-invasive fibrosis markers in biopsy-proven liver disorders. Hepatol Forum. 2025;6:16-21. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in RCA: 4] [Reference Citation Analysis (0)] |
| 13. | Najafi N, Razavi A, Jafarpour H, Raei M, Azizi Z, Davoodi L, Abdollahi A, Frouzanian M. Evaluation of hepatic injury in chronic hepatitis B and C using APRI and FIB-4 indices compared to fibroscan results. Ann Med Surg (Lond). 2024;86:3841-3846. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in RCA: 5] [Reference Citation Analysis (4)] |
| 14. | Yip TCF, Yurdaydin C. Improving prediction of hepatocellular carcinoma in chronic hepatitis B by machine learning: Productive relationship of medicine with computer science. Liver Int. 2023;43:1626-1628. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 1] [Reference Citation Analysis (0)] |
| 15. | Rui F, Xu L, Yeo YH, Xu Y, Ni W, Tan Y, Zheng Q, Tian X, Zeng QL, He Z, Qiu Y, Zhu C, Ding W, Wang J, Huang R, Xue Q, Wang X, Chen Y, Fan J, Fan Z, Ogawa E, Kwak MS, Qi X, Shi J, Wong VW, Wu C, Li J. Machine Learning-Based Models for Advanced Fibrosis and Cirrhosis Diagnosis in Chronic Hepatitis B Patients With Hepatic Steatosis. Clin Gastroenterol Hepatol. 2024;22:2250-2260.e12. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 23] [Cited by in RCA: 17] [Article Influence: 8.5] [Reference Citation Analysis (0)] |
| 16. | Xiao G, Ren H. A machine learning model to predict liver-related outcomes after the functional cure of chronic hepatitis B: Is cirrhosis driving the performance? J Hepatol. 2025;82:e141-e142. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 1] [Cited by in RCA: 2] [Article Influence: 2.0] [Reference Citation Analysis (0)] |
| 17. | Choi HSJ, Brouwer WP, Zanjir WMR, de Man RA, Feld JJ, Hansen BE, Janssen HLA, Patel K. Nonalcoholic Steatohepatitis Is Associated With Liver-Related Outcomes and All-Cause Mortality in Chronic Hepatitis B. Hepatology. 2020;71:539-548. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 185] [Cited by in RCA: 185] [Article Influence: 30.8] [Reference Citation Analysis (2)] |
| 18. | Ajuwon BI, Richardson A, Roper K, Sheel M, Audu R, Salako BL, Bojuwoye MO, Katibi IA, Lidbury BA. The development of a machine learning algorithm for early detection of viral hepatitis B infection in Nigerian patients. Sci Rep. 2023;13:3244. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 3] [Cited by in RCA: 4] [Article Influence: 1.3] [Reference Citation Analysis (0)] |
| 19. | Yoo J, Cho H, Lee DH, Cho EJ, Joo I, Jeon SK. Prognostic role of computed tomography analysis using deep learning algorithm in patients with chronic hepatitis B viral infection. Clin Mol Hepatol. 2023;29:1029-1042. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 8] [Reference Citation Analysis (0)] |
| 20. | Wai CT, Greenson JK, Fontana RJ, Kalbfleisch JD, Marrero JA, Conjeevaram HS, Lok AS. A simple noninvasive index can predict both significant fibrosis and cirrhosis in patients with chronic hepatitis C. Hepatology. 2003;38:518-526. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 3517] [Cited by in RCA: 3351] [Article Influence: 145.7] [Reference Citation Analysis (4)] |
| 21. | Sterling RK, Lissen E, Clumeck N, Sola R, Correa MC, Montaner J, S Sulkowski M, Torriani FJ, Dieterich DT, Thomas DL, Messinger D, Nelson M; APRICOT Clinical Investigators. Development of a simple noninvasive index to predict significant fibrosis in patients with HIV/HCV coinfection. Hepatology. 2006;43:1317-1325. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 4179] [Cited by in RCA: 3946] [Article Influence: 197.3] [Reference Citation Analysis (10)] |
| 22. | de Oliveira AC, El-Bacha I, Vianna MV, Parise ER. Utility and limitations of APRI and FIB4 to predict staging in a cohort of nonselected outpatients with hepatitis C. Ann Hepatol. 2016;15:326-332. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 21] [Cited by in RCA: 22] [Article Influence: 2.2] [Reference Citation Analysis (0)] |
| 23. | GBD 2019 Hepatitis B Collaborators. Global, regional, and national burden of hepatitis B, 1990-2019: a systematic analysis for the Global Burden of Disease Study 2019. Lancet Gastroenterol Hepatol. 2022;7:796-829. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 605] [Cited by in RCA: 573] [Article Influence: 143.3] [Reference Citation Analysis (4)] |
| 24. | European Association for the Study of the Liver. EASL Clinical Practice Guidelines on non-invasive tests for evaluation of liver disease severity and prognosis - 2021 update. J Hepatol. 2021;75:659-689. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 1546] [Cited by in RCA: 1415] [Article Influence: 283.0] [Reference Citation Analysis (10)] |
| 25. | Chen J, Jia J, Zhuang H, Zhang W, Yang JM, Tanwandee T, Payawal D, Hamid S, Sarin SK, Omata M, Wang G, Lau G; Apasl Viral Elimination Task Force. Assessing pricing and affordability of HBV treatment in Asia-Pacific region: a barrier to elimination. Hepatol Int. 2025;19:349-357. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 4] [Cited by in RCA: 8] [Article Influence: 8.0] [Reference Citation Analysis (0)] |
| 26. | Wang YH, Tang H, Chen EQ. Quantitative Measurement of Serum HBcrAg Can Be Used to Assess the Feasibility of Safe Discontinuation of Antiviral Therapy for Chronic Hepatitis B. Viruses. 2024;16:529. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 3] [Reference Citation Analysis (0)] |
| 27. | Vecchi A, Rossi M, Tiezzi C, Fisicaro P, Doselli S, Gabor EA, Penna A, Montali I, Ceccatelli Berti C, Reverberi V, Montali A, Fletcher SP, Degasperi E, Sambarino D, Laccabue D, Facchetti F, Schivazappa S, Loggi E, Coco B, Cavallone D, Rosselli Del Turco E, Massari M, Pedrazzi G, Missale G, Verucchi G, Andreone P, Brunetto MR, Lampertico P, Ferrari C, Boni C. HBcrAg values may predict virological and immunological responses to pegIFN-α in NUC-suppressed HBeAg-negative chronic hepatitis B. Gut. 2024;73:1737-1748. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 15] [Cited by in RCA: 14] [Article Influence: 7.0] [Reference Citation Analysis (0)] |
| 28. | Huang D, Wu D, Wang P, Wang Y, Yuan W, Hu D, Hu J, Wang Y, Tao R, Xiao F, Zhang X, Wang X, Han M, Luo X, Yan W, Ning Q. End-of-treatment HBcrAg and HBsAb levels identify durable functional cure after Peg-IFN-based therapy in patients with CHB. J Hepatol. 2022;77:42-54. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 11] [Cited by in RCA: 83] [Article Influence: 20.8] [Reference Citation Analysis (0)] |
| 29. | Fuji T, Kojima T, Kajioka H, Sakamoto M, Oka R, Katayama T, Narahara Y, Niguma T. The preoperative M2BPGi score predicts operative difficulty and the incidence of postoperative complications in laparoscopic liver resection. Surg Endosc. 2023;37:1262-1273. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 3] [Reference Citation Analysis (2)] |
| 30. | Pramono LK, Tjandrawati A, Turbawaty DK, Rostini T, Bestari MB, Haryono, Budiman D, Nugraha P. Macrophage-2-Binding Protein Glycosylation Isomer (M2BPGi) and AGAP Score as Markers of Noninvasive Test for Liver Fibrosis versus FibroScan in Chronic Hepatitis B Patients: A Retrospective Observational Study. Int J Hepatol. 2024;2024:6635625. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 3] [Cited by in RCA: 3] [Article Influence: 1.5] [Reference Citation Analysis (0)] |
| 31. | Bui HH, Nguyen ST, Phan ST, Nguyen KM, Nguyen CD. Evaluating M2BPGi as a Marker for Liver Fibrosis in Patients with Chronic Hepatitis B. Dig Dis Sci. 2023;68:4407-4417. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 12] [Cited by in RCA: 10] [Article Influence: 3.3] [Reference Citation Analysis (0)] |
| 32. | Tan Y, Zhang X. Diagnostic accuracy of FibroScan-AST (FAST) score, non-alcoholic fatty liver fibrosis score (NFS), FibroScan, and liver fibrosis index (FIB-4) for identifying fibrotic non-alcoholic steatohepatitis in patients with chronic hepatitis B with metabolic dysfunction-associated fatty liver disease. Ann Med. 2024;56:2420858. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 8] [Cited by in RCA: 9] [Article Influence: 4.5] [Reference Citation Analysis (0)] |