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

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.