Parameter-level comparison of artificial intelligence and manual cephalometric measurements: a systematic review and meta-analysis

Introduction – Prior work has largely focused on landmark-localization error and runtime; whether AI-derived parameter-level measurements differ systematically from those obtained by manual cephalometric analysis remains unclear. Objective – To compare parameter-level cephalometric measurements obtained using AI systems with manual reference methods through a systematic review and meta-analysis; secondarily, to narratively summarize any reported diagnostic metrics (κ/ICC, sensitivity/specificity, AUC) without pooling. Methods – Six databases (PubMed, Scopus, Web of Science, IEEE Xplore, EBSCO, and SciELO) were searched with no time restriction (inception to 20 September 2025) and no language limits; citation chasing was performed (Google Scholar). Eligible studies directly compared AI-based and manual cephalometric tracings. Random-effects meta-analyses were conducted only for parameter-level cephalometric measurements (primary outcomes), using standardized mean differences (Hedges’ g) with 95% confidence intervals (CIs). Statistical heterogeneity was quantified using the I2 statistic. Sensitivity/specificity/AUC were not meta-analyzed due to heterogeneous thresholds and ≤3 studies per metric; no bivariate or HSROC model was applied. Risk of bias was assessed with QUADAS-2 and certainty with GRADE. Legacy rule/knowledge-based reports were summarized narratively and not pooled. Registration: OSF (DOI: 10.17605/OSF.IO/WKMVD). Results – Twenty-two studies published between 2020 and 2025 were included. AI showed small and generally non-significant differences compared with manual methods across most evaluated parameters, although the certainty of evidence was low to moderate and heterogeneity was considerable for several outcomes. A statistically significant but small advantage was observed for the ANB angle (g = 0.28; 95% CI: 0.03–0.53; p = 0.03), with uncertain clinical relevance. Complementary legacy studies were described narratively and excluded from pooling. GRADE certainty ranged from low to moderate. Conclusions – AI-assisted cephalometric analysis produces parameter-level measurements comparable to manual tracings for most evaluated parameters; a small statistical difference for ANB was observed without clear clinical relevance. However, the certainty of evidence is limited by heterogeneity, publication bias, and lack of multicenter validation. AI may complement cephalometric workflows and save time, but it requires expert supervision and does not replace comprehensive orthodontic diagnosis or treatment planning. Funding/COI: As declared in the manuscript. Systematic Review Registration – https://doi.org/10.17605/OSF.IO/WKMVD.