Diagnostic performance of artificial intelligence in oral squamous cell carcinoma detection: a higher-order evidence synthesis through umbrella review and meta-meta-analysis

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background The integration of AI into oral squamous cell carcinoma (OSCC) diagnostics is a major milestone for digital dentistry and computational pathology, enabling faster, highly accurate detection. While numerous systematic reviews have attempted to synthesize this burgeoning evidence, their conclusions remain discordant and methodological quality varies substantially. Aim This umbrella review undertakes a rigorous synthesis of systematic review evidence to establish definitive accuracy metrics, evaluate methodological robustness, quantify evidence redundancy, and delineate critical implementation barriers. Methods A comprehensive search was conducted across MEDLINE/PubMed, Embase, Scopus, Web of Science, and Cochrane CENTRAL from inception through September 30, 2025, following PRISMA guidelines and an a priori PROSPERO-registered protocol (CRD420251157762). Systematic reviews evaluating AI for OSCC diagnosis were included. Methodological quality was appraised using AMSTAR-2, while primary study overlap was quantified via Corrected Covered Area (CCA). A meta-meta-analysis employing robust variance estimation was performed to pool diagnostic accuracy measures, accounting for statistical dependency. Results Fifteen systematic reviews synthesizing 341 primary studies were included. AMSTAR-2 assessment revealed a concerning distribution: only 3 reviews (20%) achieved high confidence, while 3 (20%) were critically low. Moderate primary study overlap was evident (CCA = 9.06%). The meta-meta-analysis, incorporating data from six reviews, yielded a pooled sensitivity of 0.90 (95% CI: 0.81–0.99; I 2 = 85%; τ 2 = 0.032). The 95% confidence interval suggested that the diagnostic performance of AI may vary substantially in future comparable reviews, reflecting differences in AI models, imaging modalities, study populations, and methodological approaches. Although the pooled estimate indicates high overall sensitivity, the wide confidence and prediction intervals indicate uncertainty regarding the magnitude and consistency of the effect, suggesting that AI performance may not be uniformly reproducible across different clinical settings Conclusions AI demonstrated compelling diagnostic accuracy for OSCC, particularly when applied to histopathological specimens. However, the evidence base is compromised by significant heterogeneity, methodological inconsistency, and a critical deficit in implementation research. This synthesis provided an authoritative evidence foundation and a strategic roadmap for researchers, clinicians, and policymakers navigating the evolving landscape of AI in oral oncology. Clinical relevance The integration of AI into routine clinical practice may assist clinicians in identifying suspicious lesions at earlier stages, improving diagnostic consistency, reducing inter-observer variability, and supporting timely referral and treatment decisions. Furthermore, AI-driven tools can facilitate large-scale screening programs, particularly in resource-constrained settings where access to specialist expertise may be limited. Although promising, successful clinical implementation requires rigorous validation, standardization of algorithms, transparency in decision-making processes, and evaluation of real-world effectiveness to ensure safe, reliable, and equitable patient care. Systematic Review Registration PROSPERO [CRD420251157762].