Although birth-dose or infant vaccination and effective antiviral therapy have made a big impact on the incidence and natural history of chronic hepatitis B (CHB), chronic hepatitis B virus (HBV) infection still affects 258 million people worldwide and remains a leading cause of cirrhosis and hepatocellular carcinoma (HCC) [1]. Hepatitis B surface antigen (HBsAg) seroclearance is a critical step in the natural history of CHB. Most patients have undetectable viremia and normal liver biochemistry after HBsAg seroclearance, and less than 5% of patients have HBsAg reversion after seroclearance [2]. This translates into a much-reduced risk of HCC and cirrhotic complications [3]. As clearance of intrahepatic covalently closed circular DNA (cccDNA) and integrated HBV DNA is unlikely to be achieved in the foreseeable future using current technologies, sustained HBsAg seroclearance with serum HBV DNA less than the lower limit of quantitation is generally accepted as the functional cure of CHB [4].

That being said, HBsAg seroclearance reduces but does not eliminate the risk of HCC, especially when HBsAg seroclearance occurs later in life and after the development of cirrhosis [5, 6]. HBsAg seroclearance significantly overlaps with occult HBV infection (OBI), which is characterized by undetectable serum HBsAg yet detectable serum and/or intrahepatic HBV DNA [7]. Isolated hepatitis B core antibody (anti-HBc) positivity is also a distinct serological pattern, defined by negative HBsAg and positive anti-HBcs. Although a few instances of OBI can be attributed to mutations in the HBsAg gene, resulting in undetectability through currently available assays, the majority of OBI cases stem from replication-competent viruses that are strongly suppressed in their replicative and transcriptional activities by the immune system of the host [8]. The ongoing low-level replication and transcription of cccDNA may lead to the risk of reactivation and silent progression of chronic liver disease [9]. Current guidelines only provide recommendations on HCC surveillance in patients with positive HBsAg. The optimal management after HBsAg seroclearance is poorly defined. With this background, this review discusses the natural history of CHB after spontaneous and treatment-related HBsAg seroclearance, factors associated with HCC development after HBsAg seroclearance, and how clinicians can apply the knowledge to select patients for surveillance.

Previous literature suggests that patients with CHB who achieve HBsAg seroclearance can still develop HCC [6]. Therefore, it is essential to assess the HCC risk in patients who have experienced HBsAg loss. Even after clearing HBsAg, patients should receive HCC surveillance if they are at risk. A more accurate risk estimation could result in a more precise recommendation for HCC surveillance.

So far, there are only a few published HCC risk scores for patients after HBsAg seroclearance [50, 56]. Existing HCC risk scores for patients with CHB, including the CU-HCC [57], REACH-B [58], PAGE-B [59], and mPAGE-B scores [60] have also been evaluated and compared with the novel HCC risk score in patients after HBsAg loss [50]. Yang et al. [50] found that their new score had a 5-year area under the receiver operating characteristic curve (AUROC) of 0.80 (0.72–0.88) and a 10-year AUROC of 0.84 (0.75–0.92), which performed better than the CU-HCC scores of 0.74 (0.66–0.82) for 5 years and 0.76 (0.68–0.84) for 10 years, and the REACH-B scores of 0.70 (0.63–0.78) for 5 years and 0.73 (0.65–0.81) for 10 years, in predicting HCC development over the respective periods (Table 2). Lee et al. [56] showed that a simple CAMP-B score (cirrhosis, age ≥ 50 years, male sex, and platelets < 1.5 × 10¹¹/L) achieved a 5-, 10- and 15-year AUROC of 0.79 (0.74–0.84), 0.77 (0.72–0.82), and 0.80 (0.74–0.85) respectively, which were comparable to that of PAGE-B and mPAGE-B scores. In CU-HCC and REACH-B scores, some weights are given to the HBV viral markers including HBV DNA and HBeAg status, which will not apply to patients after HBsAg seroclearance. The mPAGE-B score performed better than the PAGE-B score in Yang et al. study [50]. This can be because the PAGE-B score was developed in a Caucasian population while the mPAGE-B score was derived in a Korean population, which is expected to perform better in another Korean cohort from Yang et al. study [50]. The novel risk score highlighted that older age, cirrhosis, significant alcohol intake, and a family history of HCC remain important risk factors after HBsAg seroclearance. While the novel risk score performed better than existing HCC risk scores for CHB, it is important to notice that the comparison was based on internal validation. One may expect that the novel risk score can be less predictive in an external validation cohort. Also, information on significant alcohol intake and family history of HCC may be less readily available in some of the patients. In that situation, mPAGE-B score can be a good alternative as it requires only objective demographic and laboratory data and achieved an acceptably good performance in predicting 5-year and 10-year HCC development in Yang et al. study [50].

We recently explored machine learning as an alternative to traditional regression methods for HCC risk prediction. Machine learning provides a comprehensive methodology, allowing for direct parameter selection, optimizing data use, and minimizing bias. We developed models using machine learning techniques that included 46 clinical and laboratory parameters. Notably, the ridge regression model, named HCC-RS, achieved an AUROC of 0.840 and outperformed existing HCC risk scores including CU-HCC (0.672), GAG-HCC (0.745), REACH-B (0.671), PAGE-B (0.748), and REAL-B scores (0.712). These results highlight the potential of machine learning to improve HCC prediction [61]. Kim et al. [62] deployed the gradient-boosting machine algorithm on a dataset of 13,508 patients to predict HCC risk in CHB patients undergoing antiviral treatment. Impressively, this model outperformed traditional predictors, such as PAGE-B, mPAGE-B, REACH-B, and CU-HCC, achieving C-indices ranging from 0.79 to 0.81 in both Korean and Caucasian validation cohorts. This was significantly higher than the C-indices of 0.57 to 0.79 observed with the previous models [62]. In a more recent study [63], an ensemble method-based machine learning technique was introduced for HCC risk evaluation in CHB patients treated with entecavir or tenofovir disoproxil fumarate. This new approach significantly surpassed the earlier models, with an AUROC of 0.9 compared to the 0.772 of REAL-B in the training cohort. Its predictive accuracy consistently remained high with an AUROC of 0.9, in stark contrast to traditional models, such as mPAGE-B, PAGE-B, REAL-B, HCC RESCUE, and CAMD, which achieved AUROCs between 0.721 and 0.801 [63]. These results highlight the potential of machine learning to improve HCC prediction in patients with CHB.

Establishing a threshold for HCC surveillance is crucial because of the extensive resources it requires and its role in early detection and management. Its cost-effectiveness, especially for low-risk patients, is debated. HCC surveillance’s cost-effectiveness is primarily influenced by the annual incidence rate. In essence, a higher HCC incidence results in a reduced cost for each detected HCC [64]. Traditionally, the recommendation has been to initiate HCC surveillance when CHB patients have an annual risk exceeding 0.2%, and 1.5% for those diagnosed with cirrhosis [65]. Renowned organizations like AASLD [65], the Asian Pacific Association for the Study of the Liver (APASL) [66], and the European Association for the Study of the Liver (EASL) [67] have issued HCC surveillance guidelines. AASLD and EASL recommend using both abdominal ultrasonography (US) and serum alpha-fetoprotein (AFP), while APASL advises only US. Recent research pointed out that surveillance through US with AFP emerged as a cost-effective strategy in 80.1% of simulations, given a willingness-to-pay benchmark of $100,000 per quality-adjusted life-year. The study further highlighted that an HCC incidence exceeding 0.4% annually, coupled with a biannual surveillance adherence above 19.5%, is crucial for the US with AFP strategy to be more cost-effective than no surveillance at all [68].

The significance of HCC surveillance in improving early detection and management of HCC is undeniable. Guidelines from key organizations, including AASLD, APASL, and EASL, unanimously support HCC surveillance for patients with cirrhosis, specifically tailored to factors like age, ethnicity, and family history of HCC, citing substantial benefits in terms of early detection, availability of curative treatments, and enhanced overall survival [69]. Furthermore, the timing of surveillance is a crucial element of effective HCC management, with current recommendations endorsing a six-month interval [70]. This recommendation is supported by evidence linking it to heightened detection rates of early-stage HCC tumors, thereby facilitating improved treatment outcomes and prognoses. While the increased frequency of imaging has not demonstrated additional advantages, integrating modalities such as US and AFP testing has been shown to amplify the sensitivity for detecting early-stage HCC [71]. A meta-analysis highlighted US sensitivity at 94% for HCC at any stage and 63% for early-stage tumors, noting that biannual screening boosts early-stage detection sensitivity to 70% [72]. The incorporation of AFP with US has been observed to elevate early-stage HCC detection sensitivity from 32% to 65% [73]. Despite AFP’s variable efficacy, it is endorsed as a complementary tool for cirrhosis screening alongside the US, with AFP levels above 20 ng/mL achieving approximately 70% sensitivity and 90% specificity for HCC [74]. Nevertheless, EASL guidelines remain cautious regarding AFP usage due to a propensity for false positives [75]. The advent of novel biomarkers and predictive models, such as GALAD, heralds a promising future for more precise HCC detection [76].

Existing HCC risk scores are accurate for identifying patients with low HCC risk who may be exempted from HCC surveillance with a high negative predictive value of over 95% [77, 78]. However, they usually underestimate the risk of HCC among patients with high risk [78]. Thus, their utility in guiding HCC surveillance strategy remains limited. With the advancement in risk prediction using artificial intelligence and longitudinal clinical data [79], artificial intelligence-based HCC risk models in the future may better estimate the HCC risk among patients with intermediate to high risk and facilitate a risk score-based surveillance strategy. New biomarkers for HCC are also under active investigation, which may further improve the accuracy of the current risk scores [80].

Although a number of HCC risk scores have been developed for patients with CHB, few have catered for patients who have achieved HBsAg seroclearance. Furthermore, most HCC risk scores were based on baseline clinical and laboratory parameters. In the real world, however, clinicians see patients not once but repeatedly over many years. Thus, it is important to have dynamic models that cater for changes in clinical and laboratory parameters over time. Given the complexity of such data, artificial intelligence will probably be the way forward [61, 81]. As clinicians at busy clinics are unlikely to input data and apply complex models, innovative methods to draw relevant data automatically from the electronic health record and streamline the workflow will be needed.

On a separate note, a number of agents are undergoing active development with the aim of achieving functional cure of CHB [44]. Whether and how these novel therapies will impact on HCC risk is currently unknown. For example, some studies suggest that patients previously treated with peginterferon have a lower risk of HCC than those receiving NAs for CHB [82]. Therefore, it is important to examine the natural history of patients treated with new HBV drugs and determine how best to predict outcomes in such patients.

HBsAg seroclearance, also known as functional cure of CHB, reduces the risk of HCC, cirrhosis, and mortality. Nonetheless, a small proportion of patients may still develop HCC after HBsAg seroclearance. In this article, we summarized the natural history of CHB after HBsAg seroclearance, risk factors of HCC, performance of various HCC risk scores, and the way forward. To cater for this special population, separate HCC risk models are likely required, and artificial intelligence holds promise in improving prediction through handling complex and dynamic clinical data.