This study utilized data from the “China Hepatitis Prevention and Treatment Foundation - Project to Reduce Hepatocellular Carcinoma Incidence in Chronic Hepatitis B Patients” (the OASIS project, NCT04896255). This large, multicenter, prospective, real-world study enrolled CHB patients from 32 provinces across China, including treatment-naïve individuals, those previously treated with PEG-IFNα-2b, and those with nucleos(t)ide analogues (NAs) experience. Participants received either a PEG-IFNα-2b-based regimen (as monotherapy or in combination with NAs) or NAs monotherapy and are being followed for a planned 5-year period. This study was approved by the Ethics Committee of Shanghai East Hospital, Tongji University (Approval Number: [2020] Pre-review No. (157)). Written informed consent was obtained from all participants. The data were sourced from the OASIS project, with data cleaning and quality control handled by National Medical Center of Infectious Diseases (NMCID) Liver Diseases Research Group.

The specific inclusion criteria were as follows: (1) Chronic HBV infection, defined as: positivity for HBsAg for more than 6 months, or positivity for HBsAg for less than 6 months but with a liver biopsy within the past year meeting the pathological criteria for chronic hepatitis B, with other liver diseases excluded. (2) Age between 18 and 80 years, inclusive, irrespective of gender. (3)Treatment with PEG-IFNα-2b-based antiviral therapy. (4)Completion of the 48-week follow-up.

Of the 1,057 eligible CHB patients, 377 were excluded from the final analysis due to extensive missing data at baseline or the 24-week time point. Consequently, the final analytical cohort consisted of 680 patients.A flowchart detailing the study design and patient selection process is presented in Figure 1.

The clinical dataset for this study comprised demographic and clinical characteristics—including age, gender, height, weight, body mass index (BMI), cirrhosis status, and family history—of CHB patients treated with PEG-IFNα-2b. It also included serial serological parameters measured at both baseline and 24 weeks (6 months) of treatment. These parameters covered virological markers (HBV DNA, HBsAg, HBeAg, HBeAb, HBcAb), liver function indices (ALT, AST, TBIL, DBIL, ALP, GGT, ALB, GLO), and routine blood test results (PLT, NEU, Hb).The initial data processing steps included: handling missing values, applying logarithmic transformation to viral load markers (HBV DNA, HBsAg, HBeAg, HBeAb, HBcAb), and deriving kinetic variables (HBsAg_desc = baseline HBsAg level - week 24 HBsAg level; HBeAg_desc = baseline HBeAg level - week 24 HBeAg level; HBeAb_desc = baseline HBeAb level - week 24 HBeAb level; HBcAb_desc = baseline HBcAb level - week 24 HBcAb level). Categorical variables (gender, cirrhosis, family history) were factor-encoded. After excluding patients with excessive missing data, the final dataset contained 31 predictor variables. The quantitative results for HBV DNA and HBsAg were log₁₀-transformed during data preprocessing.

To identify robust predictors of treatment response, we employed five distinct feature selection algorithms. (1)Random Forest (RF): we calculated mean decrease in Gini impurity using the randomForest package to rank feature importance (Chen et al., 2024; Hosseini Sarkhosh et al., 2022). (2)Least Absolute Shrinkage and Selection Operator (LASSO): using 5-fold cross-validation via glmnet, we selected non-zero coefficients at λmin to retain features with strongest associations (Friedman et al., 2010; Simon et al., 2011). (3)Support Vector Machine-Recursive Feature Elimination (SVM-RFE): we implemented backward feature elimination with linear SVM kernel and 5-fold cross-validation (Huang et al., 2018). (4)Elastic Net: combining L1 and L2 regularization (α = 0.5), we identified features with persistent coefficients across cross-validation folds (Hughey and Butte, 2015). (5) XGBoost: we utilized gradient boosting with xgboost package, ranking features by gain importance metric which quantifies the total reduction of loss contributed by each feature. The model was configured with multi:softmax objective, maximum depth of 3, learning rate of 0.1, and 100 boosting rounds with early stopping (Chen and Guestrin, 2016; Liang et al., 2025).

Features consistently ranked highly across multiple methods were prioritized for model construction.

We trained 12 diverse classifiers using the mlr3 framework, including (1) Linear models: Logistic Regression, Linear Discriminant Analysis; (2) Tree-based models: Decision Tree, Random Forest, XGBoost; (3) Kernel methods: SVM (linear/RBF kernels); (4) Distance-based: k-Nearest Neighbors; (5)Neural networks: Single-hidden-layer Network; (6) Probabilistic: Naïve Bayes; (7) Regularized regression: GLMNet,CV GLMNet. All models were configured for probability prediction and trained on a stratified 2:1 train-test split. Hyperparameters used default values from respective mlr3learners implementations. Model performance was assessed using: Discrimination metrics: Area under ROC curve (AUC), accuracy, sensitivity, specificity, precision; F1-score; Calibration: Precision-Recall curves and net benefit analysis via Decision Curve Analysis (DCA); Stability: Comparative training-test performance and ranking consistency across metrics.

We implemented k-fold cross-validation during feature selection phases and held-out test set validation for final model assessment. Additionally, for optimal models, we performed SHapley Additive exPlanations (SHAP) analysis using DALEXtra to compute Shapley values for feature contribution quantification. Besides, We also performed univariate ROC analysis of single features and compared with ML models. This comprehensive framework ensured robust feature selection, validated model performance, and facilitated clinical interpretation of predictive factors for interferon response.

Based on HBsAg seroclearance status at week 48, patients were categorized into “Seroclearance” and “Non-Seroclearance” groups. Continuous variables with a normal distribution are presented as mean (standard deviation, SD) and compared using the independent samples t-test. Non-normally distributed variables are expressed as median (interquartile range, IQR) and compared using the Mann-Whitney U test. Categorical variables are summarized as percentages and compared using the chi-square test. All statistical analyses were performed with R software (version 4.5.1).The significance level (α) was set at 0.05, and a result with P < 0.05 was considered statistically significant.

Of the 680 CHB patients included in this study, 165 (24.3%) were classified into the seroclearance group and 515 (75.7%) into the non-seroclearance group. The data preprocessing pipeline began with the enumeration of missing values across all observations. Subsequent analysis was confined to clinical data from baseline (0W) and the 6-month (24W) time points. After excluding patients with a substantial amount of missing data, the final dataset comprised the following 31 predictor variables: Age, Gender, BMI, Liver Cirrhosis (LC), Family history, HBsAg (0W), HBsAg (24W), HBsAg_desc, HBeAg(0W), HBeAg(24W), HBeAg_desc, HBeAb (0W), HBeAb (24W), HBeAb_desc, HBcAb (0W), HBcAb (24W), HBcAb_desc, ALT (0W), ALT (24W), AST (0W), AST (24W), GGT (0W), GGT (24W), TBIL (0W), TBIL (24W), DBIL (0W), DBIL (24W), ALP (0W), ALP (24W), and PLT (0W),PLT (24W) (Supplementary Table S1).

The baseline characteristics of two groups are presented in Table 1. Differences between groups were assessed using the Chi-square test and the Mann-Whitney U test, with a statistical significance level set at α = 0.05. Significant baseline differences (all P < 0.05) between groups were observed in: LC (4.2% vs. 9.7%,P = 0.041), Family history (22.4% vs. 31.8%,P = 0.027) HBV DNA (median 0.00 vs. 1.62 log₁₀ IU/mL, P = 0.034), HBsAg (median 1.86 vs. 3.08 log₁₀ IU/mL, P < 0.001), HBeAg(median 0.40 vs. 0.52 COI, P < 0.001), HBeAb (median 0.03 vs. 0.18 COI, P = 0.001), HBcAb (median 3.29 vs. 3.28 COI, P = 0.012), ALT (median 26.70 vs. 31.00 U/L, P = 0.004), AST (median 23.70 vs. 27.00 U/L, P = 0.006), GGT (median 19.00 vs. 24.00 U/L, P = 0.001), ALB (median 46.00 vs. 45.10 g/L, P < 0.001), and Hb(median 148.00 vs. 153.00 g/L, P = 0.005). The laboratory parameters of two groups after 24 weeks of PEG-IFNα-2b treatment are presented in Table 2. Statistical comparisons were performed using the Chi-square test and Mann-Whitney U test, with the significance level set at α = 0.05. The analysis revealed statistically significant differences (all P < 0.05) in the following variables at week 24:HBsAg (median 0.00 vs. 2.47 log₁₀ IU/mL, P < 0.001), HBeAg(median 0.37 vs. 0.45 COI, P < 0.001), HBeAb (median 0.04 vs. 0.26 COI, P < 0.001), and Hb(median 131.00 vs. 135.00 g/L, P = 0.006).

Using the mlr3 framework in R, we implemented a 2:1 random split of patients into training (n = 453) and validation (n = 227) sets, followed by 500 bootstrap iterations to enhance the robustness of model assessment.Comparative analysis of baseline and 24-week treatment characteristics between the training and validation sets (Tables 3, 4) showed that, aside from family history (35.2% vs. 26.7%, P = 0.027), ALT at week 24 (median 48.80 vs. 44.00 U/L, P = 0.023), AST at week 24 (median 46.20 vs. 42.20 U/L, P = 0.048), and GGT at week 24 (median 55.00 vs. 48.00 U/L, P = 0.013), no other indicators exhibited significant differences (P > 0.05). This indicates that the baseline characteristics of the two sets were overall well-balanced, supporting their use for model training and validation.

To identify robust predictors of response to interferon-based therapy, this study employed five distinct machine learning-based feature selection algorithms: Random Forest: The results identified the following ten top-ranking features: HBsAg (24W), HBsAg (0W), HBsAg_desc, ALT (24W), GGT (24W), DBIL (24W), HBcAb_desc, HBeAb (24W), HBeAg_desc, and HBcAb (0W), as detailed in Supplementary Table S2. LASSO Regression: The final model retained the following variables: HBsAg (24W), HBeAb (24W), ALT (24W), TBIL (24W), ALP (24W), Age, and Gender, as detailed in Supplementary Table S3. SVM-RFE: The results identified the following top features: HBsAg (24W), HBsAg (0W), HBsAg_desc, HBeAg_desc, HBcAb (0W), HBeAg(24W), GGT (0W), HBeAb (24W),HBeAb_desc, and HBeAg(0W), as presented in Supplementary Table S4. Elastic Net: The algorithm identified the following key variables: HBsAg (0W), HBsAg (24W), and HBsAg_desc, as detailed in Supplementary Table S5. XGBoost: The top 10 variables were: HBsAg (24W), HBsAg (0W), HBsAg_desc, ALT (24W), DBIL (24W), GGT (24W), HBeAb (24W), ALP (24W), HBcAb (0W), and TBIL (24W), as detailed in Supplementary Table S6.

To enhance model generalizability and clinical interpretability, we defined the optimal predictor set as variables selected by at least two out of the five feature selection algorithms, which yielded 11 key predictors (Supplementary Table S7). This consensus approach is visually summarized in a Venn diagram (Supplementary Figure S1). The final predictors were: ALP (24W), ALT (24W), DBIL (24W), GGT (0W), HBcAb (0W), HBeAb (24W), HBeAg_desc, HBsAg (0W), HBsAg_desc, HBsAg (24W), and TBIL (24W). Collectively, this set captures baseline status, on-treatment levels at mid-therapy, and early kinetic changes, thereby providing multi-dimensional predictive information.

Based on the selected set of 11 key predictors, we constructed 12 distinct machine learning models, including Logistic Regression, LDA, Decision Tree, Random Forest, XGBoost, SVM (RBF), SVM (Linear), k-Nearest Neighbors, Neural Network, Naive Bayes, GLMNet, and CV GLMNet. The performance of all 12 models was rigorously evaluated on both the training and validation sets.The evaluation metrics included the Area Under the Curve (AUC), accuracy, sensitivity, specificity, precision, and F1-score. Results for each metric are reported as the mean values derived from 500 effective bootstrap samples (Supplementary Table S8). The Receiver Operating Characteristic (ROC) curve was used to assess the discriminative ability of the predictive models. These curves were plotted with 1-specificity on the x-axis and sensitivity on the y-axis, illustrating the performance for both the training and validation sets.In the training set, all models except Neural Network (AUC = 0.685 ± 0.168) demonstrated strong performance (AUC >0.85), as shown in Figure 2A. A comprehensive analysis of the ROC curves from the test set (Figure 2B) and the multi-metric radar chart (Figure 2E) identified three top-performing models: Random Forest (AUC = 0.915 ± 0.020), XGBoost (AUC = 0.891 ± 0.026), and CV GLMNet (AUC = 0.862 ± 0.029) (Figures 2C,D). In contrast, models such as k-Nearest Neighbors and Neural Network showed marked performance degradation on the validation set, with their ROC curves deviating substantially from the upper-left corner, indicating poorer generalizability.

Among the 12 machine learning models evaluated, Random Forest, XGBoost, and CV GLMNet consistently outperformed the others. The discriminative performance of Random Forest on the validation set was as follows: AUC = 0.915 ± 0.020, accuracy = 0.887 ± 0.019, precision = 0.853 ± 0.062, sensitivity = 0.707 ± 0.066, specificity = 0.954 ± 0.021, and F1-score = 0.770 ± 0.042. XGBoost achieved: AUC = 0.891 ± 0.026, accuracy = 0.878 ± 0.020, precision = 0.820 ± 0.064, sensitivity = 0.705 ± 0.063, specificity = 0.942 ± 0.023, and F1-score = 0.755 ± 0.043. CV GLMNet yielded: AUC = 0.862 ± 0.029, accuracy = 0.836 ± 0.063, precision = 0.783 ± 0.089, sensitivity = 0.567 ± 0.271, specificity = 0.940 ± 0.041, and F1-score = 0.724 ± 0.061 (Figures 3A,B).The Precision-Recall (PR) curves, plotted for both training and validation sets, demonstrated that CV GLMNet, Random Forest, and XGBoost all maintained high and stable predictive performance. Their PR curves were positioned close to the upper-right corner and exhibited narrow 95% confidence intervals, indicating a strong and reliable balance between precision and recall (Figures 4A–D).Notably, the Random Forest model achieved the highest AUC on the validation set (Figures 3A,B) and showed a balanced performance across accuracy, sensitivity, specificity, and F1-score (Figure 2F). Its performance on the validation set was comparable to that on the training set, with no significant overfitting observed. This demonstrates superior generalizability, establishing Random Forest as the optimal model for predicting HBsAg seroclearance.

This study evaluated the clinical utility of the twelve machine learning models using Decision Curve Analysis (DCA), with 95% confidence intervals calculated to reflect the reproducibility of the results. As summarized in Table 5, the threshold corresponding to the maximum net benefit was 0.01 for all models. The Random Forest model demonstrated the highest Average Net Benefit Index (ANDI) value of 0.547 and the widest range of positive threshold probabilities (0.01–0.99). These results indicate its high clinical application value, supporting its use in guiding decisions on whether to continue PEG-IFNα-2b therapy in CHB patients.Decision curves were plotted to illustrate the relationship between decision thresholds (ranging from 0.00 to 1.00) and net benefit (ranging from −0.2–0.6) across the machine learning models. Over a wide range of threshold probabilities, the net benefit of the three top-performing models—Random Forest, XGBoost, and CV GLMNet—was consistently superior to that of the other models and remained higher than the “treat-all” or “treat-none” strategies, demonstrating robust clinical discriminative ability. Among them, the Random Forest model exhibited greater adaptability to threshold variations, as evidenced by its smoother curve and stable performance across different intervention scenarios. This suggests that Random Forest is particularly well-suited for balancing benefits and risks in clinical decision-making regarding antiviral therapy for hepatitis B (Figures 5A,B).

To quantify the contribution of each predictor to the performance of the Random Forest model, we employed both the Permutation Importance method and SHAP analysis to assess variable importance (Figures 5C,D). The Permutation Importance results revealed significant differences in the importance of the 11 predictive features. HBsAg (24W) was identified as the most influential feature for model predictions, indicating that shuffling its values led to the greatest decrease in the model’s AUC. This establishes it as a core variable for maintaining the model’s discriminative ability. The following five features also provided substantial support for model performance: HBeAb (24W), DBIL (24W), GGT (0W), HBsAg_desc, and HBsAg (0W).Based on the SHAP analysis, using the shapviz package in R to compute Shapley values for features, a bar plot was generated to illustrate the contribution of the top 10 features to the model’s predictions in the Random Forest model. The ranking of feature contributions was as follows: HBsAg (24W) > HBeAb (24W) > HBsAg_desc > HBsAg (0W) > GGT (0W) > ALT (24W) > HBeAg_desc > DBIL (24W) > HBcAb (0W) > ALP (24W). Notably, HBcAb (0W) and ALP (24W) exhibited a negative predictive effect on HBsAg seroclearance, while the remaining variables showed positive predictive effects. Integrating the findings from both analytical methods, we conclude that HBsAg (24W), HBeAb (24W), GGT (0W), HBsAg_desc, and HBsAg (0W) are the key variables in the Random Forest model.

To compare the predictive performance between the machine learning models and individual variables, we conducted a univariate analysis of the 11 predictor variables on the test set, as detailed in Supplementary Table S9. ROC curves were plotted for the top ten individual features (Figures 6A,B). HBsAg (24W) was identified as the single most predictive feature, achieving an AUC of 0.866 ± 0.030, accuracy of 0.850 ± 0.025, precision of 0.723 ± 0.071, sensitivity of 0.736 ± 0.052, specificity of 0.893 ± 0.031, and an F1-score of 0.727 ± 0.046 on the validation set. Other valuable single features, in descending order of AUC, were HBsAg (0W) (AUC = 0.757 ± 0.032) and HBsAg_desc (AUC = 0.685 ± 0.036) (Figures 6C,D). In contrast, most liver function indicators—such as GGT (0W), TBIL (24W), ALP (24W), ALT (24W), and DBIL (24W)—demonstrated low predictive utility (AUC <0.6), indicating their limited value when used alone.

The univariate analysis in the test set was ranked by AUC in descending order (Figure 7A). Precision–Recall (PR) curves of the top 10 individual variables are plotted with Recall on the x-axis and Precision on the y-axis, illustrating the precision of each variable at different recall levels (Figure 7B). The PR curve for HBsAg (24W) was positioned closest to the upper-right corner, confirming its status as the best-performing single-variable predictor. However, this curve also exhibited noticeable fluctuations, reflecting its sensitivity to variations in a single feature. A sharp decline in precision was frequently observed at medium recall levels (0.4–0.6). In contrast, the Random Forest model demonstrated markedly superior and stable performance: its precision remained consistently high (mostly above 0.75) across the entire recall range (0–1) and was maintained at a high level even at high recall values (0.7–0.8), resulting in a much smoother and more robust curve. The Random Forest model, by integrating multiple predictive variables, accurately identifies patients with a high likelihood of achieving HBsAg seroclearance. This precise stratification helps prevent two critical clinical scenarios: firstly, avoiding the premature discontinuation of therapy in potential responders due to misclassification as “non-responders,” which could otherwise lead to a missed cure opportunity; and secondly, preventing the continuation of ineffective treatment in true non-responders misclassified as “responders” thereby reducing unnecessary PEG-IFNα-2b-related side effects and economic burdens (Table 6).

In this study, following data cleaning and the exclusion of patients with excessive missing data, the final dataset comprised 31 predictor variables spanning demographic and clinical characteristics, virological markers, liver function indices, and routine blood test parameters. These variables were subsequently processed using five distinct machine learning feature selection methods, which identified 11 optimal predictors. These 11 key variables were used to construct 12 machine learning models. Among them, the Random Forest model demonstrated the best performance (AUC = 0.915 ± 0.020) and surpassed the predictive capability of any single variable. Through Permutation Importance and SHAP analysis, the following variables were identified as the most influential within the Random Forest model: HBsAg (24W), HBeAb (24W), GGT (0W), HBsAg_desc, and HBsAg (0W).

HBsAg (24W) emerged as the most pivotal predictor in this study, achieving an AUC of 0.866 in the univariate analysis on the validation set and consistently ranking as the primary feature across multiple machine learning models. The quantitative HBsAg level serves as a key surrogate marker for the transcriptional activity of hepatitis B virus cccDNA. Both a lower baseline HBsAg level and a lower level at week 24 are strongly associated with a higher probability of HBsAg seroclearance at week 48 of interferon-based therapy (Tan et al., 2025; Wen et al., 2024). Furthermore, the magnitude of HBsAg decline or its absolute value at week 24 has been well-established as a robust predictor of response to interferon treatment (Jiang et al., 2024). Furthermore, incorporating the dynamic change in HBsAg (log₁₀IU/mL) from baseline to week 24 further enhanced the model’s performance, underscoring that kinetic indicators provide a more reflective measure of treatment response trends than single time-point measurements (Reijnders et al., 2011).

During antiviral therapy, seroconversion of HBeAg at week 24—defined as the loss of HBeAg accompanied by the appearance of HBeAb—serves as a powerful predictor of favorable long-term outcomes. This serological change signifies the effective activation of the patient’s immune system, enabling it to better recognize and clear HBV-infected hepatocytes. A robust and effective cellular immune response is a prerequisite for achieving HBsAg seroclearance (Liaw et al., 2012).

This study identified a significant association between baseline GGT levels and HBsAg seroclearance. In prior research, low baseline GGT has been established as an independent predictor of sustained virological response in chronic hepatitis C treatment. A lower level of GGT indicates a significant reduction in liver inflammation and bile stasis, suggesting a relative preservation of liver cell function, which may better support the antiviral immune response induced by the combination of PEG-IFN and ribavirin (Akuta et al., 2007; Dogan et al., 2014). Furthermore, serum GGT has been proposed as a potential biomarker for predicting the activation of antiviral immunity and HBeAg seroconversion in HBeAg-positive chronic hepatitis B patients receiving nucleos(t)ide analogue (NAs) therapy (Huang, 2015). However, the role of GGT levels in predicting outcomes specifically for CHB patients treated with PEG-IFNα-2b requires further investigation and validation.

In this study, the Random Forest (RF) model demonstrated superior performance. RF, an ensemble learning algorithm introduced by Breiman in 2001, operates on the core principle of “integrating votes from multiple decision trees” to model and predict complex data (Breiman, 2001). Its advantages lie in its capability to handle non-linear relationships and feature interactions, making it particularly suitable for modeling the complex associations between serological markers and treatment response (Jiang et al., 2009). Furthermore, its robustness to outliers and missing values reduces the impact of clinical data quality on model performance (Kokla et al., 2019). Permutation Importance, a model-agnostic tool for assessing feature significance, evaluates the importance of a variable by randomly shuffling its values in the validation dataset and observing the resultant decline in model performance metrics (e.g., accuracy, AUC). A greater decrease in performance indicates a higher importance of the feature to the model’s predictions. This method does not rely on the internal structure or parameters of any specific model, directly quantifying the feature’s contribution to the actual predictive performance (Fisher et al., 2019). SHAP values quantify the contribution of each feature to individual predictions, thereby assisting clinicians in understanding the model’s decision-making process and addressing the challenge of deploying traditional “black-box” models in clinical practice. By interpreting feature contributions through SHAP analysis, the model’s clinical interpretability is significantly enhanced (Lundberg and Lee, 2017; Shu et al., 2025).The combination of Permutation Importance and SHAP analysis provides complementary insights—from global feature importance to local explanations—consistently identifying HBsAg (24W) as the core predictor, with HBeAb (24W), HBsAg_desc, HBsAg (0W), and GGT (0W) as important supporting variables.For future clinical translation, a user-friendly predictive tool (such as a nomogram or a mobile application) could be developed to transform the Random Forest model into an accessible decision-support instrument for clinicians. Furthermore, prospective interventional studies are warranted to validate whether “RF model-guided therapy for CHB” can ultimately improve the rate of HBsAg seroclearance and reduce overall treatment costs.

The Random Forest model developed in this study can predict the likelihood of eventual HBsAg seroclearance in CHB patients as early as 24 weeks after initiating PEG-IFNα-2b therapy. Its clinical application includes several key aspects: For patients predicted as “non-achievers,” clinicians can promptly adjust the treatment strategy (e.g., switching to nucleos(t)ide analogues), thereby avoiding ineffective therapy and unnecessary drug-related adverse events. For those predicted as “achievers”, the positive prediction can reinforce treatment confidence and improve adherence. Furthermore, since all required variables (such as HBsAg and GGT) are routinely measured in standard clinical practice, the model can be implemented without additional testing, facilitating its widespread adoption.

To facilitate the clinical translation of our predictive model, we have outlined a clear roadmap for subsequent research. First, we will precisely calibrate the decision thresholds for predicted probabilities using a prospective cohort to establish a risk stratification framework categorizing patients as “high-, intermediate-, and low-probability responders.” Based on this, a user-friendly clinical decision support tool will be developed to enable instant input of patient indicators and automatic output of stratification-based recommendations. Ultimately, a multicenter randomized controlled trial will be conducted to validate the effectiveness of this decision-making workflow, comparing the HBsAg seroclearance rates and medical costs between the “model-guided strategy” group and the “standard management” group. This systematic effort aims to transition our model from prediction to clinical practice.