Abdulrahman Al Mopti, Abdulsalam Alqahtani, Ali H D Alshehri
Objectives: 18F-FDG PET/CT is central to head and neck cancer (HNC) care, yet most radiomic models use PET or CT alone. We quantified the discrimination of fused PET/CT models and what fusion adds. Methods: Following PRISMA 2020 and a registered protocol (PROSPERO CRD420261319855), seven sources were searched, supplemented by citation and proceedings searching. Reports sharing patients were grouped into independent cohort lineages, which served as the unit of analysis, and every synthesis included one analysis per lineage. The estimand was each report's within-study difference between its fused model and its own comparator. Two confirmatory hypotheses were specified in the statistical analysis plan, with a Bonferroni threshold of 0.025. Random-effects models used REML with Hartung-Knapp; risk of bias used PROBAST+AI, reporting completeness METRICS v1.0 and diagnostic accuracy QUADAS-3. Results: 100 studies were included, arising from 31 independent cohort lineages. Fused models exceeded PET-only models by 0.042 (95% CI 0.014 to 0.070) on the concordance index (6 lineages; p = 0.0115), meeting the confirmatory threshold, and CT-only models by 0.037 (0.002 to 0.072; p = 0.0429), which did not. Pooled absolute discrimination was 0.711 (95% CI 0.659 to 0.763). Diagnostic evidence was reported by task and not pooled. Only 25 studies could contribute to any pooled estimate, chiefly because 71 of 100 fused estimates lacked a recoverable measure of precision; all were at high risk of bias in the analysis domain. Conclusions: Fusing PET and CT radiomic information improves survival discrimination in HNC, most securely against the PET-only baseline. The advantage over CT-only models was smaller and did not meet the multiplicity-adjusted criterion, and none was established for diagnostic classification. Incremental benefit depends on the comparator, the cohort and the endpoint; it does not follow from fusion itself.