Precision stratification for hepatitis B functional cure: Biomarkers, machine learning and the combination-therapy era

Abstract

A functional cure for chronic hepatitis B remains uncommon despite effective viral suppression. Phase 3 bepirovirsen trials reported functional cure in 19%-20% of selected noncirrhotic patients who were receiving stable nucleos(t)ide analog therapy and had baseline hepatitis B surface antigen levels above 100 IU/mL and up to 3000 IU/mL. A lower baseline antigen level is associated with response across several regimens, but heterogeneous eligibility criteria and endpoints preclude attributing efficacy to selection alone. This narrative minireview examines serum biomarkers, selected prediction models, treatment-related harm, and residual hepatocellular carcinoma (HCC) risk after antigen loss. It distinguishes sustained off-treatment functional cure from antigen loss during continued background therapy, and separates pretreatment selection from on-treatment response prediction and post-clearance risk assessment. Existing models provide useful prognostic information; however, no integrated biomarker or machine learning model has been validated to allocate curative regimens. Observational evidence identifies substantial variation in HCC risk after antigen loss but does not establish the safety of surveillance withdrawal. Current society guidelines should govern clinical monitoring, whereas prospective studies should assess risk-based refinement. Diagnostic biomarker performance across genotypes does not establish transferable prediction of cure; external validation, calibration, assay standardization, and demonstrated clinical benefit remain prerequisites for implementation.

Key Words: Hepatitis B virus; Functional cure; Hepatitis B surface antigen; Biomarkers; Machine learning; Hepatocellular carcinoma

Core Tip: A low baseline hepatitis B surface antigen level is associated with response to several investigational regimens, but it should not be treated as a universal treatment-selection threshold. This minireview distinguishes functional cure from other antigen endpoints, separates baseline selection from on-treatment prediction, and appraises model validation, safety, and external validity. No integrated model is available for allocating curative regimens. Hepatocellular carcinoma risk persists after antigen loss; current guideline-based surveillance should continue where indicated, whereas prospective studies should evaluate risk-based refinement.

INTRODUCTION

Chronic hepatitis B virus (HBV) infection remains one of the most prevalent viral infections worldwide, and modeling based on Global Burden of Disease data indicates that current trajectories will fall short of the 2030 elimination target[1]. Nucleos(t)ide analog (NA) therapy has changed the natural history of disease. It suppresses viral replication in most treated patients and lowers, but does not abolish, the incidence of hepatocellular carcinoma (HCC)[2]. The residual risk is not trivial. In a European cohort of 1644 patients treated with entecavir or tenofovir, of whom 903 were followed beyond ten years, the cumulative HCC incidence reached 10.9% at 10 years and 13.2% at 15 years; the incidence rate fell from 1.25 to 0.55 per 100 person-years after year ten but did not approach zero[2].

NA therapy inhibits reverse transcription but does not eradicate covalently closed circular DNA (cccDNA) or integrated HBV DNA, both of which can sustain hepatitis B surface antigen (HBsAg) production. In the same cohort, cumulative HBsAg loss with continued therapy reached 8.3% at 10 years and 14.3% at 15 years[2]; therefore, many patients require prolonged treatment. Attempts to achieve a finite treatment course through NA withdrawal carry risks. In a prospective study of 98 non-cirrhotic, hepatitis B e antigen (HBeAg)-negative patients assigned to stop or continue therapy, HBsAg clearance occurred in 10.6% of those who stopped therapy; cumulative virological relapse by week 144 reached 83% after tenofovir withdrawal and 56% after entecavir withdrawal. Six of the 66 patients who discontinued therapy developed severe alanine aminotransferase (ALT) flares, two required hospitalization, and one died[3].

The 2022 American Association for the Study of Liver Diseases (AASLD) and European Association for the Study of the Liver (EASL) treatment endpoint conference defined functional cure as sustained HBsAg loss with serum HBV DNA below the lower limit of quantification 24 weeks after treatment discontinuation, with or without antibody to HBsAg (anti-HBs) seroconversion[4]. This review uses the definition of “off-treatment”. HBsAg loss, HBsAg reduction, eligibility to stop NA, and immune responses were described separately when the studies did not establish the full endpoint. Observational data associate HBsAg loss with improved clinical outcomes[5]; however, this does not make every antigen endpoint equivalent to a durable functional cure.

Small interfering RNA (siRNA), antisense oligonucleotide (ASO) therapeutics, capsid assembly modulators (CAMs), entry inhibitors, therapeutic vaccines, and immunomodulators are increasingly being evaluated in clinical testing[6]. Lower baseline HBsAg levels are consistently associated with response, which makes patient selection an important determinant of trial outcomes and a barrier to broad applicability. However, these associations cannot establish whether selection is more limiting than therapeutic potency. Regimen activity, prior therapy, eligibility criteria, toxicity, and follow-up duration may all contribute to the observed outcomes. In regions with a high infection burden, the potential benefit of a finite regimen must be weighed against accessibility and the effectiveness of the established NA therapy.

This review connects therapeutic evidence with specific decisions in which a biomarker or prediction model might inform pretreatment selection, treatment adaptation, and monitoring after antigen loss. It examines selected clinically relevant models rather than a comprehensive inventory of machine learning (ML) studies. The resulting framework is intended to guide research and evidence appraisal; however, it is not a validated regimen-allocation algorithm.

METHODOLOGY

This structured narrative review was updated using a literature search conducted on September 14, 2026. Ten PubMed searches used Medical Subject Headings and title/abstract terms for functional cure, investigational therapies, treatment harm, biomarkers, prediction models, NA withdrawal, post-HBsAg loss outcomes, genotype and regional applicability, and societal guidance. Publication dates were restricted to January 1, 2000, through September 14, 2026, without a database language filter. The searches returned 6401 records in total, corresponding to 5233 distinct PubMed identifiers (PMIDs). A separate search of the Cochrane Database of Systematic Reviews returned eight records; five matched PubMed records, and three were additional. Thus, 5236 distinct database records underwent title-level review. The existing bibliography was used for separate reference checking. The Supplementary material and its accompanying record files report the exact queries, source-specific counts, deduplication, and selection decisions.

Study selection was purposive for narrative synthesis. Primary English-language clinical and translational reports, current guidance, and informative evidence synthesis were considered. Pivotal randomized comparisons, negative studies, serious harm, direct model validation, and underrepresented populations were prioritized for abstract appraisal and primary-source checking. Other potentially relevant titles were recorded as not prioritized rather than counted as full-text exclusions. Preprints and conference-only reports were not used to support efficacy claims or clinical recommendations. Records were matched by PMID, with digital object identifiers (DOIs) and titles checked for duplicate reports; shared identifiers alone did not justify merging distinct publications. Scopus returned 4281 records, but its exports could not be retrieved, and these records were not screened. EMBASE was inaccessible, and Web of Science and a separate MEDLINE interface were not searched. Results from the Cochrane Central Register of Controlled Trials (CENTRAL) were not included in the screened dataset.

The appraisal considered the study design, randomization, comparator, population selection, outcome definition, denominator, follow-up, and validation settings. Within-trial randomized comparisons were given greater weight than cross-trial patterns when evaluating treatment effects. Predictor timing, number of events, missing data, validation, calibration, and clinical thresholds were examined for the prediction models. The review used no registered protocol, duplicate independent screening, formal risk-of-bias instruments, certainty-of-evidence grading, data reanalysis, or quantitative pooling. Artificial intelligence (AI) assistance is reported in the AI contribution statement: ChatGPT (OpenAI; ChatGPT 5.6) assisted with literature searching, language editing, grammatical correction, and figure design. All AI-assisted content was reviewed and verified by the author, who takes full responsibility for the integrity, accuracy, and originality of the final work. Numerical claims were checked against identifiable publication records and accessible primary text. The ledger distinguishes between the titles considered, abstracts appraised, and primary documents checked. This workflow does not constitute a systematic review.

CURATIVE PIPELINE AND THE LIMITS OF SINGLE-AGENT THERAPY

Any account of why curative regimens succeed or fail must begin with the origin of the circulating HBsAg. Two independent templates give rise to circulating HBsAg. cccDNA persists as episomal minichromosomes in the hepatocyte nucleus and directs the transcription of pregenomic RNA and pre-S/S messenger RNA (mRNA). Separately, fragments of HBV DNA integrate into the host, where they are transcribed into truncated S mRNA that yields surface antigens but no replication-competent virus. The clinical significance of this distinction was established when RNA interference produced far smaller reductions in HBsAg among HBeAg-negative than HBeAg-positive patients because in the former, the dominant antigen source was integrated DNA whose transcripts lacked the target sequence[7]. Figure 1 summarizes the pathways and points at which each therapeutic class intervenes.

Figure 1 Sources of circulating hepatitis B surface antigen and selected therapeutic targets. Covalently closed circular DNA produces pregenomic RNA and preS/S messenger RNA. Integrated hepatitis B virus DNA can also produce surface antigen, generally without supporting a complete replication cycle. RNA-silencing activity depends on retention of the compound-specific target sequence. Entry inhibitors target sodium taurocholate cotransporting polypeptide; small interfering RNA, antisense oligonucleotides, capsid assembly modulators and immune therapies act at the indicated sites. Solid arrows represent simplified pathways; dashed arrows indicate therapeutic targets[4,6,7,18]. HBV: Hepatitis B virus; NTCP: Sodium taurocholate cotransporting polypeptide; ASOs: Antisense oligonucleotides; siRNA: Small interfering RNA; HBsAg: Hepatitis B surface antigen.

Imdusiran, an N-acetylgalactosamine-conjugated siRNA, was evaluated in a phase 1a/b study involving 18 healthy participants and 65 participants with chronic HBV infection. Safety was the primary endpoint, and a total of 83 patients should not be interpreted as an efficacy denominator for functional cure[8]. In a longer follow-up of 53 recipients of ARC-520 or JNJ-3989, 1.9% achieved HBsAg seroclearance and 32.1% had HBsAg < 100 IU/mL at the last follow-up (mean 52.5 months), compared with 0% for both endpoints among five placebo recipients[9]. These results describe antigen outcomes after finite siRNA exposure and do not demonstrate an off-all-treatment cure for every responder.

PIRANGA analyzed 159 participants receiving suppressive NA therapy who were randomized to receive xalnesiran with or without an immunomodulator or NA alone. HBsAg loss 24 weeks after investigational treatment occurred in 7% and 3% of the xalnesiran monotherapy groups, 12% of the xalnesiran plus ruzotolimod group, 23% of the xalnesiran plus pegylated interferon (PegIFN) alfa-2a group, and 0% of the NA-alone group[10]. HBsAg loss occurred only among participants with screening HBsAg levels < 1000 IU/mL; however, this subgroup observation is not a prospectively validated universal cutoff. REEF-1 randomized 470 patients: 18/94 (19%) receiving JNJ-3989 200 mg with NA met the NA-stopping criteria at week 48, compared with 1/45 (2%) of the control participants; in the JNJ-3989 plus bersacapavir plus NA group, 8/94 (9%) did so[11]. However, meeting the criteria to stop NA is not equivalent to functional cure. In an earlier phase 2 elebsiran study, 11/84 participants lost HBsAg at some point, and six sustained loss 24 weeks after investigational treatment, with background NA maintained[12].

In the ENSURE trial, 55 virally suppressed participants were randomized to receive 200 mg elebsiran plus PegIFN, elebsiran 100 mg plus PegIFN, or PegIFN alone. HBsAg loss at week 72 (24 weeks after the end of investigational treatment) occurred in 4/19 (21.1%), 6/18 (33.3%), and 1/18 (5.6%) patients in the elebsiran 200 mg plus PegIFN, elebsiran 100 mg plus PegIFN, and PegIFN-alone groups, respectively; NA continued through week 72[13]. In a separate nonrandomized cohort, participants who had previously responded to BRII-179 with anti-HBs (anti-HBs responders) had HBsAg loss in 8/19, compared with 1/12 nonresponders. This supports further study of immunological profiling but does not prove that vaccination caused subsequent differences.

CAMs suppress replication through different mechanisms. In 73 virologically suppressed patients, vebicorvir added to NA did not improve HBsAg changes at week 24 compared to the placebo, despite greater suppression of replication markers[14]. Therapeutic vaccination requires careful interpretation of endpoints. In an immunological study of BRII-179 combined with siRNA, with or without interferon, sustained HBsAg loss was not achieved despite persistent anti-HBs responses in some participants[15]. This finding distinguishes immunogenicity from antigen loss; it does not establish that immune priming improves the response to subsequent combination treatment.

In the B-Clear trial, 457 participants were randomized across four bepirovirsen schedules; the 24-week 300-mg regimen achieved a sustained combined HBsAg/HBV DNA endpoint of 9%-10%, reported separately by background NA status[16]. B-Well 1 and B-Well 2 subsequently studied noncirrhotic patients receiving stable NA therapy who had baseline HBsAg levels above 100 IU/mL and up to 3000 IU/mL. Functional cure at week 72 occurred in 127/650 (20%) and 106/570 (19%) patients assigned to bepirovirsen in B-Well 1 and B-Well 2, respectively, compared with 0/328 and 0/286 patients assigned to placebo[17]. Bepirovirsen was administered for 24 weeks, with eligible participants stopping NA at week 48, and the primary endpoint was assessed 24 weeks later. The trials demonstrated efficacy within this selected population; however, because the antigen range was an entry criterion, they did not isolate the causal contribution of biomarker selection or establish efficacy above 3000 IU/mL. Mechanistic studies suggest additional intrahepatic innate immune activation through Toll-like receptor 8[18]. In the B-Together trial, bepirovirsen followed by PegIFN achieved the combined antigen/DNA endpoint in 5/55 (9%) and 8/53 (15%) patients in the two treatment groups, respectively; all responders had baseline HBsAg levels ≤ 3000 IU/mL, and background NA continued[19].

In a selgantolimod trial of 48 participants, one participant in the active treatment group met the primary endpoint of at least a 1-log10 HBsAg reduction at week 24; HBsAg loss occurred in 2/39 treated participants through week 48, with NA continued[20]. NA withdrawal is considered a separate intervention. The STOP-NUC trial randomized 166 HBeAg-negative patients and analyzed 158 (79 per group); sustained HBsAg loss by week 96 occurred in 8/79 (10.1%) patients who stopped NA and in 0/79 patients who continued it. In the subgroup with baseline HBsAg levels < 1000 IU/mL, 7/25 (28%) patients lost HBsAg[21]. Neither this outcome nor that of selgantolimod supported an unadjusted efficacy ranking across strategies. Table 1 presents the study design, treatment background, denominators, endpoints, and timing.

Table 1 Selected therapeutic studies and their distinct endpoints.

Study and design	Population and treatment background	Endpoint and assessment time	Results and denominator	Evidence maturity and interpretation
Imdusiran (AB-729-001)[8]. Phase 1a/b; randomized healthy part; open-label CHB parts	83 dosed: 18 healthy (12 active/6 placebo) and 65 CHB (22 single-dose/43 repeat-dose). Non-cirrhotic CHB; HBsAg ≥ 250 IU/mL; NA continued where prescribed	Primary: Safety. Healthy follow-up to day 29; CHB safety follow-up to 48 weeks, or 36 months in eligible NA discontinuers. Repeat dosing up to 48 weeks	No functional-cure n/N reported in the verified abstract. All 83 comprised the safety population	Early safety/antiviral study; total n includes healthy participants. Do not treat absence of a reported cure rate as zero efficacy
ARC-520/JNJ-3989 follow-up[9]. Prospective single-centre follow-up of earlier trials	53 siRNA + NA recipients (15 ARC-520/38 JNJ-3989) and 5 prior placebo recipients; 83% previously on NA; 34% HBeAg-positive	HBsAg seroclearance or HBsAg < 100 IU/mL at last follow-up (mean 52.5 months). Exact seroclearance threshold: NR; no off-all-treatment DNA composite verified	Seroclearance 1/53 (1.9%); HBsAg < 100 IU/mL 17/53 (32.1%); both 0/5 in placebo recipients	Long-term antigen durability with a small control group; not a demonstrated functional-cure comparison after withdrawal of all therapy
Xalnesiran ± immunomodulator (PIRANGA)[10]. Phase 2 randomized open-label adaptive platform	159 NA-suppressed CHB participants: Xalnesiran 100 mg (30), 200 mg (30), 200 mg + ruzotolimod (34), 200 mg + PegIFN (30), NA alone (35). NA stopping conditional	Primary: HBsAg < 0.05 IU/mL at 24 weeks after 48-week investigational treatment (week 72). HBV DNA not part of this primary endpoint	2/30 (7%), 1/30 (3%), 4/34 (12%), 7/30 (23%), and 0/35, respectively. Loss observed only with screening HBsAg < 1000 IU/mL	HBsAg loss is distinct from cure off all therapy. Small arms; exploratory baseline-HBsAg association does not establish a treatment-selection rule
JNJ-3989 ± bersacapavir (REEF-1)[11]. Phase 2b double-blind randomized trial	470 randomized across six NA-containing arms; treated/untreated and HBeAg-positive/negative strata. Reported efficacy uses modified ITT denominators	Primary: NA-stopping eligibility at week 48: HBsAg < 10 IU/mL, HBV DNA < LLOQ (20 IU/mL), HBeAg-negative and ALT < 3 × ULN	JNJ-3989 200 mg + NA: 18/94 (19%; 90%CI 13-27%); triple regimen: 8/94 (9%; 4%-15%); NA control: 1/45 (2%; 0%-10%)	Eligibility to stop NA is not functional cure. The 19% result did not involve bersacapavir; triple therapy contained JNJ-3989 100 mg
Elebsiran ± PegIFN[12]. Open-label phase 2; six cohorts	84 non-cirrhotic, NA-treated adults; entry HBsAg > 50 IU/mL and HBV DNA < 90 IU/mL. Cohort sizes 15/15/18/18/13/5; 12-48-week regimens	Primary: Safety. Secondary: HBsAg loss at any time and sustained through 24 weeks after investigational treatment. Exact HBsAg-loss threshold: NR; background NA continued	Any loss: 0/15, 1/15, 1/18, 5/18, 4/13, 0/5. Sustained loss: 0/15, 1/15, 0/18, 3/18, 2/13, 0/5 (6/84 overall)	Exploratory cohorts with unequal regimens; pooled 6/84 is not an estimate for one regimen and does not establish cure after stopping NA
Vebicorvir[14]. Randomized phase 2 placebo-controlled trial	73 non-cirrhotic NA-suppressed participants: 47 HBeAg-positive and 26 HBeAg-negative. Vebicorvir 300 mg daily or placebo added for 24 weeks	Primary: Change in HBsAg or HBeAg from baseline to week 24; continuous antigen endpoint, not an HBsAg/HBV DNA clearance composite	No between-group difference in antigen change; improved suppression of replication markers. Functional-cure n/N not reported in the verified abstract	Demonstrates antiviral activity without a demonstrated antigen-clearance advantage in this population; no inference about comparative cure efficacy
Elebsiran + BRII-179 ± interferon[15]. Phase 2a open-label randomized parallel-group study	90 CHB participants: Elebsiran alone (10), plus vaccine without interferon (41), or plus vaccine with interferon (39). Nine 4-weekly elebsiran doses	Virologic and immunologic follow-up to week 72. Sustained HBsAg loss and anti-HBs persistence evaluated; exact HBsAg-loss threshold: NR	No sustained HBsAg loss (0/90). Anti-HBs ≥ 100 IU/L persisted in approximately 40% for ≥ 32 weeks after combination treatment	Immunogenicity is not functional cure. The absence of sustained clearance in this regimen does not exclude a benefit from other vaccine-containing sequences
Bepirovirsen (B-Clear)[16]. Phase 2b randomized investigator-unblinded trial	457 ITT across four schedules: 227 on NA and 230 not on NA. For 300 mg weekly × 24 weeks: 68 on NA and 70 not on NA; background NA maintained	Primary: HBsAg < 0.05 IU/mL and HBV DNA < 20 IU/mL at every visit for 24 weeks after planned bepirovirsen end, without new antiviral rescue; week 48 for the 24-week regimen	300 mg × 24 weeks: 6/68 (9%) on NA and 7/70 (10%) not on NA	Original primary endpoint; separate from modified analyses allowing isolated blips. On-NA responses do not establish cure after stopping all therapy
Bepirovirsen (B-Well 1/2)[17]. Two replicate double-blind randomized phase 3 trials	1834 selected non-cirrhotic adults in the full analysis population on stable NA; entry HBsAg > 100 to 3000 IU/mL. Bepirovirsen 300 mg weekly or placebo for 24 weeks; eligible NA withdrawal at week 48	Primary: Functional cure at week 72, ≥ 24 weeks after all treatment stopped: HBsAg not detected (< 0.05 IU/mL) and HBV DNA < LLOQ (< 20 IU/mL or target not detected)	B-Well 1: 127/650 (20%) vs 0/328; B-Well 2: 106/570 (19%) vs 0/286	Replicated efficacy within an enriched population. HBsAg-based enrollment was not a randomized selection strategy; durability beyond current follow-up remains to be established
Bepirovirsen then PegIFN (B-Together)[19]. Phase 2b randomized open-label trial	108 ITT on stable NA throughout: 24-week (55) or 12-week (53) bepirovirsen, followed by up to 24-week PegIFN	Primary: HBsAg < 0.05 IU/mL and HBV DNA < 20 IU/mL at every visit for 24 weeks after planned PegIFN end, without new antiviral rescue; background NA continued	5/55 (9%) and 8/53 (15%), respectively; all responders had baseline HBsAg ≤ 3000 IU/mL	Proof of concept; no bepirovirsen-alone concurrent control. Cross-trial relapse comparisons do not prove the benefit of adding PegIFN
Selgantolimod[20]. Randomized phase 2 placebo-controlled trial	48 NA-suppressed participants (24 HBeAg-positive/24 negative): 39 active and 9 placebo. Weekly study drug for 24 weeks; NA continued through week 48	Primary: ≥ 1 log10 IU/mL HBsAg decline at week 24. Additional HBsAg loss through week 48; exact loss threshold: NR. No off-all-treatment composite	Primary endpoint 1/39 active; HBsAg loss 2/39 (5%) active vs 0/9 placebo through week 48	Small exploratory immunomodulator study; HBsAg loss on NA does not establish finite-treatment functional cure
NA withdrawal (STOP-NUC)[21]. Multicentre randomized stop-vs-continue trial	166 randomized; 158 analyzed, 79 per arm. HBeAg-negative; continuous long-term NA with HBV DNA < 172 IU/mL for ≥ 4 years	Primary: Sustained HBsAg loss to week 96. HBsAg assay LOD 0.05 IU/mL; HBV DNA assay LOD 10 IU/mL. Primary endpoint differs from the B-Well composite	8/79 (10.1%; 95%CI 4.8-19.5%) vs 0/79; P = 0.006. Baseline HBsAg < 1000 IU/mL subgroup: 7/25 (28%)	Selected withdrawal trial; 11/79 (13.9%) retreated. No cessation-related serious adverse events observed; small numbers do not exclude uncommon serious harm
Elebsiran + PegIFN (ENSURE)[13]. Partially randomized open-label phase 2 trial	55 randomized to PegIFN alone or plus elebsiran 200/100 mg; 31 previously vaccinated participants in a separate nonrandomized cohort. Non-cirrhotic, NA-suppressed; randomized entry HBsAg > 100 IU/mL to 3000 IU/mL	HBsAg < 0.05 IU/mL at end of 48-week treatment and 24 weeks later (week 72); HBV DNA assay LLOQ 10 IU/mL. NA continued until week 72; withdrawal assessed afterward	Week 72: 4/19 (21.1%) with 200 mg, 6/18 (33.3%) with 100 mg, 1/18 (5.6%) control. Prior-vaccine cohort: 8/19 anti-HBs responders vs 1/12 nonresponders	Randomized regimen comparison, but small samples and HBsAg enrichment. Prior-vaccine response is an exploratory association; endpoint does not establish cure off NA
OCTOPUS-1[22]. Open-label phase 2; randomized nivolumab schedules	37 virally suppressed adults; JNJ-3989 + NA with one (n = 18) or three (n = 19) nivolumab infusions. Background NA continued through week 72	Primary: HBsAg < 0.05 IU/mL at 24 weeks after the 24-week intervention (week 48); follow-up to week 72. HBV DNA LLOQ 20 IU/mL was a separate virological measure	Primary HBsAg loss: 0/18 and 0/19. One participant in the three-infusion arm achieved loss at 48-week follow-up	No internal arm without nivolumab; incremental benefit was assessed against an external REEF-1 cohort. No off-NA cure estimate. No serious or grade 3/4 adverse events; two transient asymptomatic hyperthyroidism cases
REEF-IT[23]. Open-label phase 2; regimen amended during study	54 untreated HBeAg-positive adults; JNJ-3989 + NA ± bersacapavir for 36-52 weeks, followed by 12-week PegIFN add-on. Most continued NA during follow-up; only one met NA-completion criteria	Primary: HBsAg < 0.05 IU/mL at 24 weeks after stopping all study treatment, without NA restart. Follow-up continued for 48 weeks. Assay threshold verified in NCT04439539	0/51 in the registered primary per-protocol set (54 enrolled). HBsAg loss at least once: 11/54 (20.4%); six maintained loss through follow-up week 48	On-treatment or transient antigen loss did not establish the off-all-treatment endpoint. Registry primary denominator is distinguished from the publication’s enrolled population. No deaths or serious adverse events; two stopped PegIFN
ALG-020572[24]. Double-blind, placebo-controlled phase 1	32 healthy participants (24 active/8 placebo); eight non-cirrhotic, HBeAg-negative, NA-suppressed CHB participants (6 active/2 placebo). CHB participants continued NA	Safety, pharmacokinetics and antiviral activity; planned up to seven 210-mg doses in CHB, with 12-week follow-up after the last dose. Trial stopped early for liver injury	ALT > 10 × ULN attributed to drug-induced liver injury in 4/6 active-treated CHB participants after 2-7 doses; one hospitalized. No functional-cure estimate	Three biopsies suggested drug-induced liver injury. All recovered without sequelae; development was terminated. Compound-specific serious toxicity must not be recast as a beneficial immune flare or generalized to all ASOs
Biomarker-guided NA withdrawal[25]. Multicentre randomized trial (2:1)	90 Asian patients in the modified ITT analysis: 60 stopped and 30 continued NA. HBsAg < 200 IU/mL, suppressed HBV DNA ≥ 12 months, and undetectable HBV RNA or HBcrAg; cirrhosis excluded	Published analysis: Cumulative HBsAg loss through 144 weeks. Qualitative HBsAg negativity; numerical loss threshold NR. Registry NCT04519359 lists a 72-week primary endpoint	Cumulative HBsAg loss 25.9% vs 4.2% (P = 0.013); raw event n/N NR. In the stop arm, 16/60 restarted NA; no serious adverse events were reported	Cumulative estimates are not raw proportions. Supports a selected withdrawal strategy, not isolated validation of RNA/HBcrAg added value or a broadly applicable safe-stopping rule

Evidence maturity is a narrative appraisal, not a formal certainty grade. Background NA continuation, eligibility and differing assessment intervals prevent direct cross-trial comparisons. HBsAg loss, antigen decline and NA-stopping eligibility are not interchangeable with functional cure after withdrawal of all treatment. NA: Nucleos(t)ide analogue; CHB: Chronic hepatitis B; HBeAg: Hepatitis B e antigen; HBsAg: Hepatitis B surface antigen; siRNA: Small interfering RNA; ASO: Antisense oligonucleotide; CAM: Capsid assembly modulator; PegIFN: Pegylated interferon; ALT: Alanine aminotransferase; ULN: Upper limit of normal; ITT: Intention to treat; LOD: Limit of detection; LLOQ: Lower limit of quantification; NR: Not retrievable in the sources checked for this revision, not evidence that the original report omitted the information.

Negative studies have highlighted the distinction between antigen decline and a cure. In the OCTOPUS-1 trial, 37 suppressed participants were randomized between two nivolumab schedules on a JNJ-3989 and NA backbone; none achieved the primary HBsAg loss endpoint at the 24-week follow-up, although one participant lost HBsAg later. As there was no concurrent group without nivolumab, the incremental benefit was assessed against external trial data[22]. The REEF-IT trial enrolled 54 HBeAg-positive participants who were untreated at enrollment. The registered primary per-protocol analysis recorded no off-all-treatment HBsAg-seroclearance responses among 51 participants at the 24-week follow-up, although 11/54 had lost HBsAg and only one met the NA-completion criteria[23]. These findings caution against inferring a durable cure based on on-treatment antigen responses.

The repeated association between lower HBsAg levels and response supports the testable enrichment hypothesis; however, the thresholds have different meanings across studies. PIRANGA and B-Together describe responder distributions, STOP-NUC reports a subgroup after withdrawal, and B-Well restricts eligibility. These findings provide prognostic information in the context of treatment. Demonstrating treatment effect prediction requires evidence that the biomarker modifies comparative treatment benefits, for example, a pre-specified biomarker-by-treatment interaction or a randomized comparison of selection strategies. A lower antigen level might reduce immune exhaustion, or it might simply require a smaller absolute decline to cross the assay threshold. Antigen sources, immune restoration, and cure are incompletely linked in humans; however, the underlying mechanisms remain unclear.

TREATMENT-RELATED HARM AND THE COST OF FINITE THERAPY

The benefits of treatment must be considered in terms of harm. Within PIRANGA, grade 3 or 4 adverse events occurred in 50% of the xalnesiran plus PegIFN group and 6% of the NA-only group[10]. Therefore, the higher HBsAg loss rate in the interferon-containing arm was accompanied by a greater adverse event burden in that trial. This observation supports an explicit safety assessment when considering interferon-containing research regimens; however, it does not establish a general efficacy-toxicity ranking across different drugs.

In the pooled B-Well safety analysis through week 72, adverse events occurred in 91% of bepirovirsen recipients and 73% of placebo recipients, and serious adverse events occurred in 7% and 4% of the patients, respectively. During treatment, grade 3 or higher events occurred in 16% and 3%; ALT elevation was the most frequent severe event in the bepirovirsen group (6%)[17]. Nausea has been reported in 46% of the selgantolimod recipients[20]. These estimates retained their original comparators and observation windows. Differences in event ascertainment, severity definitions, treatment exposure, eligibility, and follow-up between trials prevented direct comparisons of the absolute adverse event rates.

ALT elevation also requires cause-specific interpretation. In the phase 1 ALG-020572 ASO study, four of six actively treated participants with chronic hepatitis B (CHB) developed ALT levels more than 10 times the upper limit of normal; one was hospitalized, and three underwent biopsy with findings consistent with drug-induced liver injury. Dosing and development were stopped, and all affected participants recovered without clinical sequelae[24]. This small study does not estimate the risk for the ASO class, but shows why ALT flares cannot automatically be interpreted as beneficial immune activity.

NA withdrawal requires caution when it is a part of a finite treatment strategy. The prospective study described above recorded six severe ALT flares, two hospitalizations, and one death among 66 patients who discontinued NA[3]. The STOP-NUC trial reported no serious adverse events attributed to NA cessation, and 13.9% of patients required retreatment[21]. Both studies used protocolized monitoring, and differences in population, stopping criteria, retreatment rules, and follow-up may explain this discrepancy. However, these studies did not establish whether protocolized withdrawal is safer than routine care. Any future selection strategy must be evaluated together with a defined monitoring and retreatment pathway; not every investigational HBsAg-loss endpoint requires the withdrawal of background NA.

This biomarker-guided withdrawal trial provides prospective evidence in a selected setting. The modified intention-to-treat analysis included 90 participants with HBsAg levels < 200 IU/mL, undetectable HBV DNA for at least 12 months, and undetectable HBV RNA or hepatitis B core-related antigen (HBcrAg); 60 stopped and 30 continued NA. The 144-week cumulative HBsAg loss estimates were 25.9% and 4.2%, respectively; 16 participants in the withdrawal group were retreated, and no serious adverse events were reported[25]. These are cumulative estimates and not raw event fractions. The trial supports the evaluated withdrawal strategy but does not isolate the incremental value of RNA or HBcrAg over quantitative HBsAg or validate the allocation among novel curative combinations.

BIOMARKERS AND THEIR LIMITATIONS

Quantitative HBsAg analysis provides information about the antigen burden and its trajectory during treatment. GOLDEN used baseline and six-monthly HBsAg measurements in 6,792 NA-treated patients, among whom 173 HBsAg loss events occurred (61/2835 in development and 112/3957 in external validation). The five-year cumulative incidence of loss was 2.4%. Reported areas under the receiver operating characteristic curve (AUCs) were 0.981 (95%CI: 0.974-0.987) and 0.979 (95%CI: 0.974-0.983), respectively[26]. The full report includes calibration analyses and a mean lead time of 18.9 months among correctly predicted HBsAg-loss cases; the overall positive predictive value was 26.2%, illustrating why high discrimination does not itself imply a high probability of the rare outcome for every flagged patient. Repeated measurements were used to update predictions; therefore, this was a longitudinal NA treatment model, not a pretreatment test for novel-regimen selection. Predictor values must be available at each prediction origin and must precede the predicted outcome when a model is deployed or revalidated. No specific information leakage errors are identified. Large-scale antigen profiling and later treatment cohorts further illustrate the effects of disease phase and measurement timing[27,28].

HBcrAg provides information on viral transcription. An individual-participant meta-analysis assembled 10397 participants from 60 studies, and the principal diagnostic analysis included 5591 untreated patients. AUCs for detecting HBV DNA at or above 2000 and 200000 IU/mL were 0.88 (95%CI: 0.83-0.94) and 0.96 (95%CI: 0.94-0.98), using HBcrAg cutoffs of 3.6 Log U/mL and 5.3 Log U/mL. The validation sensitivity and specificity at the lower DNA threshold were 85.2% and 84.7%[29]. The diagnostic performance did not vary materially according to genotype. The correlation with HBV DNA was poor during antiviral treatment, which does not establish that HBcrAg lacks value as a marker for residual transcription. This is evidence for detecting specified viral load thresholds and not genotype-independent prediction of functional cure. Paired serum and liver studies support serum HBV RNA as a marker of intrahepatic cccDNA transcription in untreated and NA-treated patients[30], without establishing a universal treatment-selection rule.

The quantitative antibody to the hepatitis B core antigen (anti-HBc) is a host-response-associated marker, and baseline titers have been associated with HBeAg seroconversion during PegIFN and NA treatment[31]. This endpoint differs from that of HBsAg loss and functional cure. These markers may provide complementary information; however, anti-HBc is not a direct measure of global immune competence, and a combination of low HBsAg, low HBV RNA, and high anti-HBc has not been validated to allocate a curative regimen. Assay variability, disease phase, genotype distribution, and contribution of integrated DNA can alter the meaning of the measured antigen value.

The CREATE study evaluated end-of-treatment HBsAg and HBcrAg levels as predictors of response after NA cessation, a specific withdrawal setting, rather than for selecting novel combinations[32]. HBcrAg and HBV RNA also face assay access and standardization constraints. Measurements below the detection limits are left-censored and may retain prognostic information despite reduced resolution. Values obtained during NA therapy may not have the same interpretation as pretreatment measurements. Therefore, endpoint consensus and contemporary biomarker reviews regard several candidate markers as exploratory[4,33]. Combining markers is a research option, whose incremental value must be tested against a simpler model.

The incremental predictive value was context-dependent. Among 150 NA-suppressed patients receiving PegIFN add-on, combining baseline HBV RNA and HBcrAg with HBsAg increased the reported AUC for week-48 HBsAg loss from 0.785 to 0.845[34]. Conversely, in 128 untreated patients with HBeAg-negative chronic infection, single-time-point HBcrAg and circulating HBV RNA levels added no predictive value beyond quantitative HBsAg levels for spontaneous HBsAg loss[35]. These clinical settings and endpoints differ and neither result establishes a universal biomarker combination.

PREDICTION MODELS AND THEIR INTENDED USE

Nonlinear associations, interactions, and longitudinal measurements motivate flexible prediction methods; however, appropriately specified regression models can also represent these features. An advantage of ML must be demonstrated by comparison with well-developed conventional models using the same predictors, outcomes, and validation data. Pretreatment regimen selection, early response prediction, and risk assessment after HBsAg loss are different clinical tasks, and evidence for one does not validate the other.

Dong et al[36] developed a LightGBM model for 2882 inactive carriers treated with PegIFN, including 541 week-24 response events, with 2017 participants used for training and 865 for internal testing. The model used baseline and week-12 information, including the week-12 ALT/HBsAg ratio, to predict week-24 HBsAg below 0.05 IU/mL with HBV DNA below 20 IU/mL. The internal test AUC was 0.891, and the external AUC in 167 participants, including 33 events, was 0.917 (95%CI: 0.850-0.983). The study reports random forest imputation, a Brier score of 0.091, and decision curve analysis; the Brier score alone does not establish calibration. Because external performance contributed to final model selection, the external cohort cannot be regarded as a fully independent validation set. Whether all preprocessing and feature selection steps were repeated within the resampling folds was not established in this report. It is an on-treatment prediction model with a short horizon. It cannot select a regimen before therapy begins or establish a sustained cure after withdrawal.

Baseline models were also evaluated after PegIFN therapy. FC-48w, using pretreatment age and HBsAg and ALT levels, predicted HBsAg loss during 48 weeks of PegIFN add-on in 377 NA-suppressed HBeAg-negative patients. External validation in 148 patients from five other Chinese centers yielded an AUC of 0.852 (95%CI: 0.785-0.905)[37]. SHAN and FLASH-N use baseline clinical variables to predict sustained HBsAg clearance for at least 24 weeks after treatment cessation, with external AUCs of 0.861 (95%CI: 0.805-0.916) and 0.800 (95%CI: 0.716-0.884), respectively[38]. Their operational outcomes did not specify the complete antigen and DNA endpoint used to define a functional cure. These retrospective studies predicted the outcomes among PegIFN recipients; however, they did not establish comparative treatment benefits or validate the allocation between curative regimens.

PLAN-B-CURE addresses a different task: Liver-related outcomes after HBsAg seroclearance. It was developed in a Korean training cohort of 944 patients, internally validated in 1102 patients, and externally evaluated in 2741 patients from Hong Kong. The 123 events and incidence of 1.1% per person-year reported in the abstract refer to the combined Korean cohort, not to all 4787 participants[39]. Because the outcome combines HCC, hepatic decompensation, and liver-related death, the model must not be treated as an HCC-specific surveillance model. The training concordance index (C-index), AUC, and area under the precision-recall curve (AUPRC) were 0.82, 0.86, and 0.53, respectively. The reported Hosmer-Lemeshow P value above 0.05 does not by itself prove accurate calibration. The primary abstract states that the performance was reproduced in the validation cohorts, but the exact validation estimates could not be independently extracted from the primary material available for this review. Table 2 shows the extraction limitations instead of asserting that the publication omitted the validation results. The AUPRC and positive predictive value help interpret rare outcomes; rarity does not mathematically inflate the AUC or prove that the model is biased.

Table 2 Selected prediction models by intended use and validation setting.

Model/intended use	Population, events and prediction timing	Discrimination and validation	Calibration/overall accuracy	Development and missing data	Thresholds and limits for clinical use
GOLDEN[26]: Longitudinal discriminant analysis; HBsAg loss during NA therapy	China: Training 2835/61 events; geographic validation 3957/112 events; 173/6792 overall. qHBsAg every 6 months; loss < 0.05 IU/mL twice ≥ 6 months apart. Median follow-up 65.6 months; 5-year incidence 2.4%	AUROC: Training 0.981 (95%CI 0.974-0.987); validation 0.979 (0.974-0.983). Overall PPV 26.2%; mean prediction lead time 18.9 months among correctly predicted HBsAg-loss cases	Calibration plots for 5-year persistence of HBsAg in Supplementary Figure S5 of the original report[26]; slope/intercept not extracted	Nanfang Hospital training; 49 other centers validation; fivefold cross-validation. ≥ 2 qHBsAg samples ≥ 6 months apart required; explicit imputation strategy not located in examined main text	Score > 0.95 at any follow-up identified favorable trajectories; 5-year loss 17.0% training/29.1% validation. Dynamic outcome-proximal prediction differs from a baseline regimen-selection rule. Landmark evaluation is needed before treatment allocation; this audit does not establish leakage
Dong et al[36]: LightGBM; rapid on-treatment PegIFN response	Inactive HBeAg-negative Chinese patients, baseline HBsAg < 1000 IU/mL. Development n = 2882/541 events; training 2017, test 865; external 167/33 events. Baseline/week-12 inputs predict week-24 HBsAg < 0.05 IU/mL plus DNA < 20 IU/mL	AUROC training 0.902 (95%CI 0.881-0.923); internal test 0.891; external 0.917 (0.850-0.983). External precision 0.547, sensitivity 0.879, specificity 0.821	Brier score 0.091 and decision curves reported; Brier is overall predictive error, not a stand-alone calibration assessment. Calibration slope/intercept not located in examined text	70:30 stratified split; tenfold cross-validation; six selected predictors. Variables > 20% missing excluded; random-forest imputation; training-based normalization. Full nesting of preprocessing within resampling not established from report	Proposed probability < 0.3 PegIFN stopping rule; no prospective impact trial. Week-12 ALT/HBsAg ratio dominated SHAP. Week-24 response is not sustained off-treatment cure; patients already cleared by week 12 were not explicitly excluded. Model-choice discussion considers external performance, limiting a strictly untouched-test interpretation
SHAN/FLASH-N[38]: Baseline prediction of sustained HBsAg loss with PegIFN	Chinese HBeAg-negative patients: Derivation 202/94 events; overlapping advantaged subgroup 114/58. SHAN training/test 142/60; external 183, including 117 advantaged patients. Baseline predictors; HBsAg < 0.05 IU/mL sustained ≥ 24 weeks after cessation. External event counts not extracted	SHAN AUC training 0.908 (95%CI 0.860-0.955), test 0.949 (0.898-0.999), external 0.861 (0.805-0.916). FLASH-N AUC 0.920; bootstrap-corrected 0.889; external 0.800 (0.716-0.884)	Internal calibration plots and internal/external decision curves reported. External calibration plot or slope not identified in the examined body. Decision curves estimate net benefit; no impact trial	48 candidates; univariate screening, tenfold LASSO and stepwise logistic regression; bootstrap/Firth correction for FLASH-N. Key missing data, > 10 missing variables or absent follow-up excluded; MICE for remaining missingness. Full fold-wise nesting not established	SHAN uses ln(HBsAg+1), age, sex and neutrophil percentage; FLASH-N adds AFP/LDH. Retrospective, selected PegIFN recipients; 524/726 initially eligible patients excluded. No prospectively tested treatment-allocation threshold or comparative regimen selector
FC-48w[37]: Baseline Cox-derived score; HBsAg loss during PegIFN add-on	377 Chinese NA-suppressed, non-cirrhotic HBeAg-negative patients; development five centres 229/41 events, external five other centres 148/30. Baseline age, HBsAg and ALT; 48-week PegIFN plus NA. Off-treatment durability not established	AUC development 0.842 (95%CI 0.788-0.887); external 0.852 (0.785-0.905). Validation HBsAg loss: Low 2/73, intermediate 17/55, high 11/20	Calibration assessment not identified in examined main text. Risk-stratum separation does not establish calibration of individual probabilities	Univariate screening followed by forward stepwise Cox regression. 485 screened; 38 declined and 70 lost to follow-up excluded. Imputation strategy not identified in examined body. Most patients had HBsAg < 1500 IU/mL	Score = 0.008 × ALT (U/L) - 0.039 × age (years) - 0.002 × HBsAg (IU/mL). Derivation cutoffs -2.7/-1.3 target ≥ 90% sensitivity/≥ 85% specificity. Prognostic strata among PegIFN recipients; suggested treatment choices not prospectively tested against usual care or another regimen
REACH-B[44]: Cox-derived points score; HCC in untreated CHB	3584 untreated non-cirrhotic Taiwanese patients/131 HCC; external Hong Kong/Korea 1505/111 HCC. Baseline sex, age, ALT, HBeAg, HBV DNA; 3-year, 5-year, 10-year horizons	External AUROC 0.811 (95%CI 0.790-0.831), 0.796 (0.775-0.816), 0.769 (0.747-0.790) at 3/5/10 years. Later post-loss validation: 5-year AUC 0.64 (95%CI 0.52-0.75), n = 1200/23 HCC[47]	Predicted-vs-observed calibration plots, Figures 1 and 2. Correlations at 3/5/10 years 0.973/0.942/0.994; correlation alone is not equivalent to calibration slope/intercept	Independent hospital validation including 277 cirrhotic patients; separate non-cirrhotic analysis. Development median follow-up 12.0 years; validation mean 7.3 years. Missing-data strategy not located in examined main text	17-point score estimates continuous absolute risk; not developed as a post-HBsAg-loss surveillance stopping rule. Treatment and loss alter distributions/meaning of virological inputs
PAGE-B[45]: Cox-derived points score; HCC during ETV/TDF	Caucasian patients: Derivation 1325/51 HCC; external ninth-center validation 490/34 HCC. Age, sex, platelets at treatment initiation; 5-year horizon. Median follow-up 50 months overall	C-index derivation 0.82; bootstrap 0.81; external validation 0.82. Confidence intervals not provided in the examined main text. Later post-loss validation: 5-year AUC 0.62 (95%CI 0.49-0.75)[47]; cirrhosis comparative evaluation C-index 0.605 (0.568-0.644)[50]	Calibration plots for 5-year HCC-free probability in derivation and validation (Supplementary Figure 1A/B of the original report[45]); observed risk strata also assessed	Multiple imputation: 10 datasets, predictor selection in > 50%, Rubin pooling; 1000 bootstrap samples; geographic holdout	≤ 9, 10-17, ≥ 18: 5-year HCC 0%/3%/17% in derivation and 0%/4%/16% in validation. Developed during continuing antiviral treatment; a low score alone does not validate stopping surveillance after cure. Cohort overlaps the long-term PAGE-B report[2]
aMAP[46]: Cox-derived score; HCC across liver-disease etiologies	17374 participants; training 3688/95 HCC; nine validation cohorts 13686/536 HCC (sum of Table 1). Age, sex, albumin, bilirubin, platelets. Generally 5-year horizon; one cohort 3 years	Training C-index 0.82 (95%CI 0.77-0.86). CHB validation 0.82-0.87; e.g., European PAGE-B 0.82 (0.78-0.86). Cirrhotic subgroup values 0.61-0.83. Later post-loss validation: 5-year AUC 0.73 (95%CI 0.63-0.84)[47]; cirrhosis comparative evaluation C-index 0.603 (0.564-0.641)[50]	Observed-vs-predicted calibration plots for training and all validation cohorts (Figure 2); no slope/intercept extracted	Mixed treated HBV/HCV and non-viral cohorts. Available scores: 3662 training and 13324 validation; imputation method not located in examined main text	Cutoffs 50/60; validation NPV 99.3%-100% at 50, PPV 6.6%-15.7% at 60. Neither broad etiologic validation nor low observed incidence establishes a post-HBsAg-loss stopping threshold; no treatment-allocation role
PLAN-B[49]: Gradient boosting; HCC during NA treatment	Training 6051; external Korean 5817 and Caucasian PAGE-B 1640. Ten baseline parameters; HCC follow-up reported through 8 years. Event counts and complete input list not retrievable from primary sources examined	C-index training 0.79; external Korean 0.79, Caucasian 0.81. No significant advantage over modified PAGE-B in Caucasians (P = 0.42); confidence intervals not retrievable from sources examined. Later cirrhosis comparative evaluation C-index 0.638 (95%CI 0.600-0.675)[50]; independence from all original cohorts should not be assumed	Primary abstract explicitly reports a satisfactory calibration plot. Numerical calibration parameters not retrievable from sources examined	Independent geographic cohorts on ETV/TDF. Missing-data, tuning and internal-validation details not retrievable from primary sources examined	Minimal-risk group 11.2% Korean/8.8% Caucasian; < 0.5% HCC by 8 years. Not a post-seroclearance model. Ethnic labels do not establish viral genotype; between-cohort C-index comparisons do not measure calibration
PLAN-B-CURE[39]: Gradient boosting; composite liver-related outcomes after HBsAg loss	Training 944, internal validation 1102 (Korea), external 2741 (Hong Kong). Seven inputs: Age, sex, diabetes, alcohol, cirrhosis, albumin, platelets. Korean cohort: 123 composite events over median 55.2 months; 1.1%/person-year	Training C-index 0.82, AUROC 0.86, AUPRC 0.53. Primary abstract states replication in both validation cohorts; separate values/confidence intervals not retrievable from primary sources examined	Hosmer-Lemeshow P > 0.05 reported. Non-significance does not establish accurate individual risk estimates; slope/intercept not retrievable from sources examined	Six Korean centers plus Hong Kong database, seroclearance 2000-2022. Complete horizon-specific, event-split and missing-data methods not retrievable from primary sources examined	Outcome combines HCC, decompensation and liver-related death; not interchangeable with HCC risk. No verified threshold for withdrawing HCC surveillance. 123 events apply to Korea, not the total 4787. Full-text availability is an audit limitation, not evidence of deficient reporting
CAMP-B[48]: Cox-derived points score; HCC after HBsAg seroclearance	Korean development 3476/102 HCC, 24019 person-years; Hong Kong external validation 5255/85 HCC. Cirrhosis, age ≥ 50, male sex, platelets < 150000/mm3 after HBsAg loss; 5-year, 10-year, 15-year horizons	Development time-dependent AUC 0.782/0.817/0.825 at 5/10/15 years; external 0.785/0.771/0.796. AUC confidence intervals not retrievable from examined primary abstract	Calibration details not retrievable from examined primary abstract; publisher full text inaccessible. This is an access limitation, not proof of non-reporting	Retrospective development at two Korean hospitals with geographic external validation in Hong Kong. Multivariable Cox model. Missing-data and internal-validation details not retrievable from examined abstract	Points: Cirrhosis, 3; age ≥ 50 years, 2; male sex, 3; low platelets (< 150000/mm3), 1. External low/intermediate/high annual HCC incidence 0.07%/0.37%/0.90%; exact stratum cutoffs not extracted. Dedicated post-loss prediction does not establish safe surveillance withdrawal

Missing or unextracted fields describe the limits of accessible primary evidence, not proof of absent reporting. Models address different outcomes, time origins and populations and cannot be ranked directly. The later post-loss entries for REACH-B, PAGE-B and aMAP use the same 1200-patient cohort, not three independent cohorts. None supplies a validated integrated baseline regimen-allocation rule or a tested strategy for withdrawing indicated surveillance. PegIFN: Pegylated interferon; AFP: Alpha-fetoprotein; LDH: Lactate dehydrogenase; LASSO: Least absolute shrinkage and selection operator; MICE: Multiple imputation by chained equations. AUROC/AUC: Area under the receiver operating characteristic curve; AUPRC: Area under the precision-recall curve; PPV: Positive predictive value; NPV: Negative predictive value; HCC: Hepatocellular carcinoma; NA: Nucleos(t)ide analogue; ETV: Entecavir; TDF: Tenofovir disoproxil fumarate; qHBsAg: Quantitative hepatitis B surface antigen; ALT: Alanine aminotransferase; SHAP: Shapley additive explanations; HBsAg: Hepatitis B surface antigen; CHB: Chronic hepatitis B; HBeAg: Hepatitis B e antigen; HBV: Hepatitis B virus; HCV: Hepatitis C virus; C-index: Concordance index.

Society guidance and the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis statement updated for AI (TRIPOD + AI) emphasize transparent development, validation, and reporting[40,41]. Table 2 presents the purpose, timing, events, validation, calibration, and thresholds of the selected models. External discrimination should be accompanied by calibration and clinical utility assessments, including the uncertainty and consequences of false-positive and false-negative classifications. Validation pre-processing must use only the information available at the intended prediction time, and the evaluation should preserve the intended prediction origin and follow-up horizon. A prospective evaluation must determine whether model-guided care improves outcomes compared to usual care. Figure 2 shows the conceptual research framework; no integrated model combining the proposed inputs has been validated for allocating curative regimens.

Figure 2 Conceptual research framework for precision stratification. A: Candidate baseline inputs require model development, external validation and prospective utility trials; no integrated regimen selector is validated; B: Existing longitudinal quantitative hepatitis B surface antigen models during nucleos(t)ide analogue therapy and week-12 alanine aminotransferase/hepatitis B surface antigen (HBsAg) prediction during pegylated interferon (PegIFN) address different outcomes and timepoints. Baseline PegIFN models predict within-regimen outcomes; neither baseline nor on-treatment prediction establishes comparative regimen allocation or adaptation utility; C: After HBsAg loss, hepatocellular carcinoma surveillance follows guideline indications. HCC-specific risk and composite liver-related outcomes require separate validation before changing surveillance. Genotype D/E calibration and clinical utility remain research priorities[26,36,38,39,41,48,51-55]. HBV: Hepatitis B virus; qHBsAg: Quantitative hepatitis B surface antigen; HBcrAg: Hepatitis B core-related antigen; anti-HBc: Antibody to hepatitis B core antigen; HCC: Hepatocellular carcinoma; PegIFN: Pegylated interferon; HBsAg: Hepatitis B surface antigen; NA: Nucleos(t)ide analogue; ALT: Alanine aminotransferase.

A prospective evaluation has been initiated. The GIFT protocol plans to assess the GOLDEN threshold using quantitative HBsAg trajectories recorded before PegIFN add-on in 300 NA-experienced patients. Both score strata will receive PegIFN plus NA, and the primary endpoint will be any HBsAg loss during 48 weeks, irrespective of the status at week 48[42]. This is a planned prognostic validation within one regimen, and there is no evidence that model-guided treatment improves outcomes.

RESIDUAL HEPATOCELLULAR CARCINOMA RISK AFTER FUNCTIONAL CURE

In a territory-wide cohort of 13379 patients with HBsAg loss, 274 (2.0%) developed HCC after HBsAg loss. The annual incidence in patients with cirrhosis was 0.43% for those who lost HBsAg before 50 years of age and 0.62% for those who lost HBsAg at or above 50 years of age. Among patients without cirrhosis who lost HBsAg before 50 years of age, the incidence was 0.03%; no HCC occurred during a follow-up of up to 15 years among 1317 women in this subgroup[43]. These findings identify heterogeneity in the observed risk. Zero observed events do not prove zero future risk, and administrative fibrosis classification, sex, ethnicity, duration of infection, competing mortality, and time since antigen loss affect the interpretation.

Conventional HCC risk scores have been developed in different settings. REACH-B used 3584 untreated Taiwanese patients for development and 1505 patients from Hong Kong and South Korea for validation, with validation AUCs of 0.811, 0.796, and 0.769 at three, five, and ten years, and reported agreement between predicted and observed risk[44]. PAGE-B was developed in 1325 and validated in 490 NA-treated Caucasian patients, with a C-index of 0.82 in both cohorts[45]. The aMAP score was developed and validated in 17374 patients with a mixed liver disease etiology. It reported C-indices of 0.82-0.87 in CHB validation cohorts and calibration plots, with lower discrimination in some cirrhosis-only analyses[46]. These scores were developed outside the post-HBsAg-loss populations; however, subsequent direct validation is available. In 1200 patients with HBsAg seroclearance and 23 HCC events, five-year AUCs were 0.62 (95%CI: 0.49-0.75) for PAGE-B, 0.73 (95%CI: 0.63-0.84) for aMAP, and 0.64 (95%CI: 0.52-0.75) for REACH-B[47]. CAMP-B was specifically developed after HBsAg seroclearance in 3476 Korean patients with 102 HCC events and was externally validated in 5255 Hong Kong patients with 85 events. Its external five-, ten-, and fifteen-year AUCs were 0.785, 0.771, and 0.796[48]. This supports post-HBsAg-loss risk stratification but leaves the safety of withdrawing otherwise indicated surveillance unproven.

The PLAN-B gradient-boosting HCC model was developed in 6051 NA-treated patients and evaluated in Korean (n = 5817) and Caucasian (n = 1640) validation cohorts with C-indices of 0.79 and 0.81, respectively, and calibration plots[49]. The minimal-risk group comprised 11.2% and 8.8% of the validation cohorts and had < 0.5% HCC risk over 8 years. In the European cohort, the difference in discrimination between PLAN-B and modified PAGE-B was not statistically significant (P = 0.42). This does not establish an equal performance, failed calibration, or a genotype-specific explanation. However, the model was not developed for post-HBsAg loss surveillance. Table 2 distinguishes this HCC endpoint from the PLAN-B-CURE composite outcomes.

Performance can deteriorate in different clinical settings. In a later comparative evaluation of 1339 patients with cirrhosis who started potent NA therapy, the C-indices were 0.605 for PAGE-B, 0.603 for aMAP, and 0.638 for PLAN-B[50]. These findings concern a population with cirrhosis and do not prove universal model failure. Differences in case mix and predictor timing require explicit evaluation before a model can be transferred to another decision or population.

Surveillance decisions should follow current guidelines. The 2025 Joint AASLD and Infectious Diseases Society of America HBV guidelines suggest continued HCC surveillance after HBsAg loss in patients with cirrhosis, a family history of HCC, men who lose HBsAg after age 40 years, and women who lose HBsAg after age 50[51]. The EASL guidelines retain surveillance for patients with persistent HCC risk after HBsAg loss[52,53]. The 2026 Asian Pacific Association for the Study of the Liver HBV guidelines recommend lifelong ultrasound and alpha-fetoprotein surveillance every 6 months after HBsAg loss in patients with advanced fibrosis or cirrhosis[54]. Its inclusion of patients with advanced fibrosis is broader than that of the EASL HCC guidelines, which cite insufficient evidence for routine surveillance in advanced fibrosis without cirrhosis[53]. Therefore, recommendations should follow the guidelines of the relevant society and clinical setting. The AASLD HCC guidelines use approximately 0.2% annual risk for selected non-cirrhotic HBV populations and approximately 1% for cirrhosis as cost-effectiveness benchmarks; these are model-dependent population thresholds and are not validated individual stopping rules[55]. The observed 0.03% annual incidence in the younger non-cirrhotic subgroup was below the former benchmark; however, uncertainty, fibrosis misclassification, family history, hepatitis D coinfection, metabolic risk, alcohol exposure, and regional risk must be considered. Neither this cohort nor low model scores established that surveillance could be safely withdrawn. These recommendations do not require surveillance for all patients after HBsAg loss. Withdrawing otherwise indicated surveillance based on a model requires prospective validation and decision analyses comparing HCC stage at diagnosis, mortality, harms, and costs.

EXTERNAL VALIDITY AND THE GENOTYPE QUESTION

Many cure-response and risk models have been developed for East Asian and European cohorts. HBV genotype distribution varies across Africa, with genotype E being common in West Africa and genotype D being important in North Africa, including Egypt[56]. Region or ethnicity is not a substitute for individual genotyping, and a Caucasian validation cohort cannot be assumed to represent genotype D. For the proposed integrated cure-selection approach, direct outcome validation in the genotype D and E populations remains absent from the evidence reviewed.

Four questions must be addressed. Assay discrimination asks whether a biomarker identifies a measured state, such as an HBV DNA threshold; HBcrAg has relevant cross-genotype evidence for this task[29]. The biomarker-outcome association asks whether the marker is associated with a future cure in a defined treatment context. Calibration asks whether the predicted absolute probabilities agree with the observed outcomes in the target population. Clinical utility assesses whether using predictions improves decisions and outcomes. Evidence for the first question does not resolve the other three questions. Baseline risk, age, fibrosis, co-infection, treatment access, and competing mortality can affect transportability independent of the genotype.

Regional diagnostic validation has been performed for other biomarkers. In a Gambian study of 339 HBV-infected patients, Mac-2 binding protein glycosylation isomer was evaluated for fibrosis and prevalent HCC; its validation AUC for diagnosing cirrhosis was 0.89[57]. This cross-sectional diagnostic evidence provides regional information but does not validate the prediction of future functional cure, future HCC, or the proposed integrated regimen selection framework.

The difference in the comparative discrimination between the Korean and Caucasian PLAN-B validation cohorts demonstrates why performance should be examined separately for each target setting[49]. There is no evidence that calibration specifically failed or that the genotype caused this difference. Underrepresentation of African populations in health AI datasets is a broader concern[58]; however, regional implementation requires direct evaluation of the HBV model and the outcome in question.

Regional validation should record predictors at the intended decision time; ascertain genotypes where relevant; measure sufficient outcome events; and report discrimination, calibration-in-the-large, calibration slope, and clinical net benefit. Recalibration may be appropriate if baseline risk differs, but it should not be assumed to resolve the altered predictor effects or poor discrimination. In settings where HBcrAg or HBV RNA are unavailable, routinely collected quantitative HBsAg, age, platelet count, and fibrosis assessments offer feasible inputs for research. PAGE-B and aMAP illustrate the feasibility of parsimonious HCC risk assessments in their original treatment settings[45,46]; however, they do not validate a minimum-variable rule for curative regimen allocation or surveillance withdrawal.

INTEGRATION WITH GUIDELINE-BASED CARE

Current clinical care should include guideline-based treatment eligibility, fibrosis assessment, virological monitoring, and HCC surveillance[51,52,53,55]. A research model may supplement these assessments after validation for its intended use but should not override contraindications, patient preferences, or access constraints. If NA cessation is considered, eligibility, monitoring frequency, and retreatment criteria should be specified in advance, with reliable access to reviews for biochemical or virological relapse. The proposed minimal-variable models require a direct comparison with usual care, and their feasibility does not establish clinical benefit.

LIMITATIONS OF THE STUDY

This narrative synthesis had no prospectively registered protocol, duplicate independent screenings, or formal risk-of-bias assessments. Title-level coverage was broad; however, abstract and primary-source appraisals were selective. PubMed and Cochrane review coverage, English-language selection, unavailable Scopus exports, inaccessible EMBASE, and incomplete full-text access may have led to missing relevant evidence. The models were selected as examples rather than as an exhaustive inventory. Trial response and safety rates differ by population, treatment background, assay thresholds, denominators, and observation windows; cross-trial patterns remain hypothesis-generating. No pooled estimate was calculated, and the durability of the latest regimens remains uncertain. The prediction models and observational surveillance findings were not independently reanalyzed or prospectively tested.

CONTRIBUTION OF THIS REVIEW

Recent reviews have described various combination therapies and emerging biomarkers[6,33]. This minireview connects these topics to the timing and purpose of clinical prediction, distinguishes assay discrimination from cure prediction, and presents treatment harm and efficacy. The proposed contribution is a decision-focused research framework whose assumptions are explicit and testable.

The framework separates three tasks that should not be conflated: Selecting participants before treatment, estimating their response after treatment begins, and assessing liver-related risk after HBsAg loss. It also separates HCC-specific outcomes from composite liver-related outcomes and explains why neither high discrimination nor a low observed incidence alone establishes a safe change in surveillance.

For genotype D and E settings, the practical research priority is to establish outcome-specific transportability and feasibility of the required assays. This perspective supports parsimonious validation studies without implying that currently available scores can allocate curative therapy or justify surveillance withdrawal.

CONCLUSION

Patient selection is an important determinant of response and limits the applicability of current functional cure evidence; however, it has not been shown to be more limiting than therapeutic potency. The phase 3 bepirovirsen results apply to selected non-cirrhotic patients receiving stable NA therapy and with baseline HBsAg levels above 100 IU/mL and up to 3000 IU/mL. Associations between lower antigen levels and response across other regimens support prospective enrichment research but not a universal treatment threshold.

The next steps are assay standardization, explicit assessment of treatment benefits and harm, and validation of prediction models for the decision and population in which they will be used. Baseline models predict outcomes within particular treatment settings, whereas longitudinal and week-12 models require measurements obtained after treatment initiation. Neither method establishes a comparative selection of curative regimens. No integrated biomarkers or ML algorithms are available for this allocation task in routine care.

HBsAg loss reduces, but does not abolish, the risk of HCC. Observational cohorts support further risk stratification, whereas guideline-based surveillance remains appropriate, as indicated. Prospective outcome studies and decision-analytic evidence are required before model-based withdrawal of otherwise indicated surveillance is recommended. Validation in populations with genotype D and E must address cure outcomes, calibration, and clinical utility separately from diagnostic biomarker performance.