Antiangiogenic class and trial population are confounded in first-line immunotherapy for hepatocellular carcinoma: a Bayesian network meta-analysis of nine randomised trials

Background First-line therapy for advanced hepatocellular carcinoma (HCC) combines an immune checkpoint inhibitor (ICI) with one of two antiangiogenic backbones: an anti-VEGF monoclonal antibody or a VEGFR-targeted tyrosine kinase inhibitor (TKI). No head-to-head trial exists, and successive network meta-analyses have ranked the two classes and reported class-level probabilities as estimates of a drug-class property. Objective To estimate the class comparison and, separately, to test whether a class effect is identifiable from the aggregate trial data on which such rankings are based. Methods Nine randomised trials (5,333 patients) were synthesised in a Bayesian random-effects network anchored on sorafenib. Because each treatment appears in a single trial, between-study heterogeneity is not identifiable from the data; it was therefore specified using the external empirical predictive distributions of Turner et al., and every efficacy estimate reported here is conditional on that externally sourced prior. Outcomes were overall survival (OS), progression-free survival (PFS), objective response (ORR) and grade ≥ 3 adverse events. Five effect modifiers were pre-specified; their joint distribution across trials was examined before interpreting any class-level estimate. Results Three regimens improved OS over sorafenib: sintilimab plus bevacizumab (HR 0.57, 95% credible interval 0.35–0.92), finotonlimab plus bevacizumab (0.60, 0.37–0.99) and camrelizumab plus rivoceranib (0.64, 0.41–0.99). The posterior probability favouring the anti-VEGF-antibody class was 75.0% for OS, 74.8% for ORR and 45.3% for PFS. These estimates are not interpretable as class effects. Treatment class, hepatitis B virus prevalence, macrovascular invasion and China-only enrolment were mutually collinear across trials (pairwise r = 0.81–0.91; HBV–China r = 0.87), so a survival advantage attributed to drug class is equally consistent with one attributable to the populations in which each class was tested. With nine trials, trial-level meta-regression is underpowered to separate the independent contributions of viral aetiology and geographic enrolment and is reported as hypothesis-generating only; no reweighting or prior specification separates them either, and sensitivity analyses moved the class probability only between 73.9% and 85.5% without altering this. The contrasts that do survive are those estimated within randomised comparisons: grade ≥ 3 hand-foot syndrome occurred with every TKI regimen (2–12%) and with none of the four anti-VEGF-antibody regimens, against 6.9–12% in the sorafenib arms of those same trials. Conclusions Class rankings in first-line HCC describe the evidence base rather than a property of the drugs. The collinearity of treatment class with viral aetiology and geographic region makes class attribution unidentifiable from aggregate data, irrespective of model choice, and this limitation applies to the published rankings as much as to ours. On GRADE-NMA the class-level indirect comparison was rated very low certainty, downgraded for unresolvable confounding by trial population. Separating drug class from trial population will require individual-patient-data analysis or head-to-head randomisation; until then, within-trial contrasts remain the defensible basis for class-level claims.