Real-world outcomes of atezolizumab-bevacizumab and durvalumab-tremelimumab in unresectable hepatocellular carcinoma: a systematic review with covariate-adjusted meta-regression and exploratory machine learning analysis.

Background/purpose: Atezolizumab-bevacizumab (AB) and durvalumab-tremelimumab (DT) are standard first-line therapies for unresectable hepatocellular carcinoma (uHCC), but their comparative effectiveness remains uncertain. Methods: We included 76 observational reports comprising 10,243 patients. Report-level median survival was summarized descriptively, whereas objective response rate (ORR) and disease control rate (DCR) were synthesized using random-effects models. Covariate-adjusted meta-regression was the principal adjusted analysis, and adverse events were summarized descriptively. An exploratory prediction analysis using leakage-free nested cross-validation was performed, with an S-learner as the primary model and T-learners only as supporting analyses. Results: The evidence base comprised 60 AB reports (8,928 patients) and 16 DT reports (1,315 patients). For median overall survival (mOS), the sample size-weighted means were 17.7 months for AB and 16.5 months for DT; the corresponding medians of report-level medians were 17.1 and 13.8 months, respectively. These survival summaries were descriptive and were not formally compared. Random-effects pooled ORR was 26.6% for AB and 20.8% for DT, and pooled DCR was 71.4% and 55.4%, respectively. In covariate-adjusted meta-regression, the AB-DT differences were +3.57 months for mOS (95% CI, -0.38 to +7.52; P = 0.076), +1.96 months for median progression-free survival (mPFS) (+0.91 to +3.02; P < 0.001), +2.95 percentage points for ORR (-2.65 to +8.54; P = 0.302), and +16.15 points for DCR (+10.44 to +21.86; P < 0.001). The mPFS association was not reproduced in first line-predominant cohorts (+1.35 months, -0.33 to +3.04; P = 0.116) and attenuated after the simultaneous exclusion of four reports with unresolved overlap (+1.69 months, +0.31 to +3.07; P = 0.016). Conclusion: In available aggregate real-world data, AB cohorts reported higher median PFS and disease control than DT cohorts, whereas OS could not be reliably compared. Marked differences in treatment line, age, calendar period, and covariate support, together with the small DT evidence base, preclude reliable comparative-effectiveness inference. Systematic Review Registration: https://www.crd.york.ac.uk/PROSPERO/view/CRD420251141735, identifier CRD420251141735.