Chronic hepatitis B (CHB) remains a major global health burden despite decades of antiviral therapy and vaccination efforts.¹ Hepatitis B surface antigen (HBsAg) seroclearance is widely accepted as a functional cure and the optimal therapeutic endpoint in international guidelines because complete viral eradication is currently unattainable due to the persistence of covalently closed circular DNA.^2,3 Achieving a functional cure is associated with a substantial reduction in the risk of hepatocellular carcinoma (HCC), hepatic decompensation, and liver- related mortality.^4-7 On the other hand, accumulating evidence suggests that the risk of liver-related complications does not disappear completely after HBsAg loss.⁸ Long-term cohort studies have revealed a persistent annual incidence of HCC after HBsAg seroclearance that remains clinically meaningful. For example, the annual HCC incidence in a Korean cohort of patients with HBsAg seroclearance was reported at 0.86%, exceeding the generally accepted cost-effectiveness threshold of 0.2% per year for HCC surveillance in patients with CHB.^9,10 With the development of novel curative therapies and increasing numbers of patients achieving functional cure,¹¹ determining which patients remain at risk has become a critical unmet need. Risk stratification in a post-cure setting is a critical clinical imperative. Therefore, functional cure should be viewed not as the termination of the disease but as a transition into a distinct clinical state characterized by residual hepatocarcinogenic potential.

The continued risk of liver-related complications, particularly HCC and hepatic decompensation, after HBsAg seroclearance indicates the biological consequences of a chronic infection. Even after the viral transcriptional activity becomes minimal, integrated HBV DNA and accumulated genetic alterations within hepatocytes continue to promote carcinogenesis.^12,13 Fibrotic remodeling and microenvironmental changes within the liver may remain irreversible in some patients, affecting the long-term outcomes.¹⁴ Interestingly, different liver-related outcomes may follow different trajectories after a cure. In a large, territory-wide cohort of more than 9,700 patients with HBsAg seroclearance, the risk of hepatic decompensation tends to decline over time, whereas the risk of HCC persists for years.⁸ In this study, the annual incidence of HCC remained stable over time (0.20% during the first seven years vs. 0.19% during years 8–12; p=0.898), while the annual incidence of hepatic decompensation decreased from 0.26% to 0.12% after seven years (p=0.009). Furthermore, HBsAg loss for more than seven years was independently associated with a 45% reduced risk of hepatic decompensation, but not with a lower risk of HCC. Therefore, post-cure hepatocarcinogenesis is no longer directly linked to viral replication, but instead to host-related factors, such as fibrosis burden, age-related susceptibility, and metabolic injury.^15-17 Accordingly, the clinical phenotype of CHB shifts toward a host-driven chronic liver condition after cure, in which the underlying liver vulnerability, rather than the original viral etiology, becomes the principal determinant of the long-term prognosis.

With the accumulation of outcome data from patients who achieved HBsAg seroclearance, regression-based prediction models were developed to estimate the residual HCC risk. A Korean cohort study involving 831 patients who achieved HBsAg seroclearance reported an annual HCC incidence of 0.86%, and identified the age at seroclearance, cirrhosis, family history of HCC, and alcohol consumption as the independent predictors of HCC.⁹ A simple risk score based on these four variables showed good discrimination, with a C-index of 0.804 and time-dependent area under the receiver operating characteristic curve (AUROC) values of 0.799, 0.835, and 0.817 at five, 10, and 15 years, respectively. Another model aimed to improve the clinical applicability by incorporating objective, routinely available parameters. The CAMP-B score, derived from an international multicenter cohort of over 8,700 patients, uses cirrhosis, age ≥50 years, male sex, and platelet count to stratify the HCC risk after HBsAg loss.¹⁸ In the validation cohort, the annual incidence of HCC ranged from 0.07% in the low-risk group to 0.90% in the high-risk group, with time-dependent AUROCs approaching 0.80 across the long-term follow-up.

These conventional statistics-based models have provided important conceptual and clinical advances. Their major strength lies in their simplicity. They rely on readily available clinical parameters and can be easily implemented in routine practice without requiring specialized testing. Nevertheless, this simplicity also reflects inherent limitations. These models are based on static baseline variables, assume linear relationships between the predictors and outcomes, and are typically optimized for predicting a single endpoint such as HCC. Therefore, they may insufficiently capture the complex, nonlinear interplay between the fibrosis burden, aging, metabolic injury, and dynamic liver function that determines the long-term outcomes after cure. As a result, although conventional scores offer useful population-level risk stratification, they remain limited in their ability to support truly individualized post-cure surveillance strategies.

The persistent and multifactorial nature of the post-cure risk has driven the application of machine learning methods capable of modeling complex interactions among variables. Recently, a machine learning-based model (PLAN-B-CURE) to predict liver-related outcomes, including HCC, hepatic decompensation, and liver-related mortality, was developed and validated in a large, multinational cohort involving 4,787 patients who achieved HBsAg seroclearance.¹⁵ One hundred and twenty-three liver-related events were observed during a median follow-up of 55.2 months, corresponding to a 1.1% per per-son-year incidence rate. The final model, constructed using a gradient boosting algorithm and seven routinely available clinical variables―age, sex, diabetes, alcohol intake, cirrhosis, albumin level, and platelet count―revealed improved predictive performance compared to the conventional HCC risk scores. The model achieved a C-index of 0.82 and an AUROC of 0.86, whereas traditional models in the same cohort achieved AUROCs ranging from 0.62 to 0.72. Importantly, the PLAN-B-CURE model was designed to predict a composite outcome comprising HCC, hepatic decompensation, and liver-related mortality, rather than a single endpoint, reflecting the clinical reality that post-cure prognosis encompasses multiple competing events. Risk stratification based on the model identified distinct groups with significantly different cumulative incidences of liver-related outcomes over time, supporting its ability to provide an individualized prognostic estimation.

Artificial intelligence-based prediction models have been applied across CHB management. Beyond post-cure risk stratification, machine learning can help predict a functional cure during pegylated interferon therapy, enabling more informed treatment decisions before seroclearance is achieved.¹⁹ In another multinational study, a machine learning model incorporating routine clinical and laboratory variables revealed cross-ethnic generalizability in predicting the HCC risk across Korean and Caucasian populations with CHB.²⁰ More recently, a CT-based imaging biomarker model further extended this approach by integrating radiological features, showing that predictive inputs do not need to be limited to conventional laboratory parameters.²¹ These models differ in their clinical setting and their target outcomes and input features, reflecting the diverse applications of machine learning across the CHB disease spectrum. Importantly, machine learning models estimate the risk on a continuous, individualized scale rather than assigning patients to predefined categorical groups, thereby providing a practical foundation for personalized surveillance strategies in the era of increasing functional cure.

The current practice guidelines recommend continued surveillance in patients with cirrhosis after seroclearance.^2,3 For non-cirrhotic patients, however, guidance remains limited and based largely on heterogeneous risk factors without standardized thresholds. The American Association for the Study of Liver Diseases (AASLD) guidelines suggest surveillance for selected high-risk non-cirrhotic individuals, including those with a family history of HCC and those who achieve seroclearance at an older age (men ≥40 years and women ≥50 years). In contrast, the European Association for the Study of the Liver (EASL) guideline broadly recommends surveillance in “at-risk” patients, with risk stratification based on clinical risk factors rather than a single standardized algorithm.^2,3 As a result, surveillance is often applied uniformly despite the substantial heterogeneity in risk. Accurate risk prediction models may help address this gap by enabling individualized surveillance strategies. High-risk patients may require close surveillance similar to cirrhotic populations, whereas very-low-risk individuals may be candidates for less intensive monitoring. Nevertheless, direct evidence supporting reduced-intensity surveillance in low-risk patients after seroclearance is limited, and such recommendations are currently based on low-level evidence extrapolated from cost-effectiveness modeling rather than prospective clinical data. Such risk-adaptive strategies can help improve cost-effectiveness and reduce the patient burden while maintaining clinical safety. On the other hand, prediction models must undergo prospective validation in real-world clinical settings before such approaches can be implemented in routine practice. In addition, formal cost-effectiveness analyses will be important to support the clinical implementation of these risk-adaptive surveillance strategies because the cost-effectiveness of surveillance depends on multiple factors, including surveillance modality, healthcare resources, and regional disease epidemiology.^10,22,23 Therefore, decisions regarding the surveillance intensity should incorporate the individualized biological risk and local clinical context.

The future of post-cure management lies in precision medicine. Advances in multi-omics platforms, biomarker discovery, and artificial intelligence analytics will enable more refined risk predictions. Beyond binary risk classification, a multi-tier risk stratification framework incorporating intermediate strata would better reflect the continuum of a clinical risk. Dynamic prediction models incorporating longitudinal data may improve accuracy beyond single-time-point assessments, enabling risk estimates that evolve with changing clinical status.²⁴ Ultimately, CHB management is transitioning from uniform surveillance to adaptive monitoring, wherein the surveillance intensity responds to changing biological risk.

Nevertheless, several challenges must be addressed before machine learning-based prediction can be fully integrated into clinical practice. The model performance may be affected by overfitting, particularly when developed using limited or homogeneous datasets. Thus, external validation across diverse populations and healthcare settings is essential for ensuring generalizability. In addition, the “black box” nature of machine learning algorithms may limit clinical interpretability, underscoring the need for transparent modeling approaches and explainable artificial intelligence techniques. The SHAP (SHapley Additive exPlanations) values, as incorporated in the PLAN-B-CURE model, represent one such approach by quantifying the contribution of each variable to individual risk predictions, thereby enhancing interpretability without sacrificing the predictive performance. Nevertheless, a fundamental trade-off exists between the model complexity and clinical transparency. Highly complex models may achieve superior discrimination but remain difficult to interpret and implement at the bedside, whereas simpler models sacrifice some predictive accuracy in favor of clinical utility.²⁵ Achieving the right balance between these competing priorities will be essential for the successful translation of machine learning-based tools into routine clinical practice. Furthermore, existing prediction models, including PLAN-B-CURE, have been developed in cohorts that encompassed spontaneous and nucleos(t)ide analog-induced seroclearance, but they did not include patients achieving a functional cure through novel therapeutic agents such as capsid assembly modulators or RNA-targeting therapies. Whether these models apply to such treatment-induced functional cure populations remains to be determined and warrants a prospective evaluation as novel therapies enter clinical practice. Future research should focus on improving the predictive accuracy and enhancing the robustness, interpretability, and real-world applicability of machine learning-based risk stratification tools.

Despite functional cures substantially reducing the risk of disease progression in CHB, they do not eliminate the possibility of future liver-related complications. A persistent hepatocarcinogenic potential driven by host and liver-related factors necessitates continued risk assessments after HBsAg seroclearance. Conventional prediction models are simple and clinically accessible but have limited capacity to support individualized surveillance strategies. Emerging machine learning approaches now enable more precise prognostication by capturing the complex host-driven determinants of long-term outcomes. As the number of patients achieving a functional cure continues to grow, the focus of CHB care is shifting from viral suppression toward managing the residual risk after cure. In this context, risk-adaptive surveillance informed by accurate stratification may improve the clinical outcomes and cost-effectiveness. Ultimately, the transition from uniform monitoring to risk-adaptive surveillance will be essential to advancing CHB management from viral control to precision hepatology.