Artificial intelligence in cervical cancer screening and triage: a role-stratified systematic review and bivariate meta-analysis.

PURPOSE OF REVIEW: Artificial intelligence (AI) tools for cervical cancer screening have proliferated, but modality-pooled accuracy estimates conflate clinically distinct uses of AI. We re-examine this evidence base through a role-stratified bivariate meta-analysis to clarify where AI is ready for clinical translation and where gaps remain. RECENT FINDINGS: Of 97 eligible studies published between 2019 and 2026, 47 with reconstructible 2 × 2 data were pooled using a bivariate Reitsma model stratified by clinical role. Diagnostic assistance during colposcopy showed the most mature evidence ( k  = 21; sensitivity 0.908, specificity 0.844; HSROC AUC 0.94). Primary AI-cytology screening showed high sensitivity but unstable specificity ( k  = 10; 0.934/0.701 [0.460-0.865]; AUC 0.92). Triage of hrHPV-positive women was the smallest and weakest pool ( k  = 4; sensitivity 0.805, 95% CI lower bound 0.624 - below the 90% safety threshold commonly cited for HPV-positive triage). External validation was reported in 10.9% of studies and 49.1% originated from China. Deeks' funnel asymmetry was borderline for screening ( P  = 0.082) and significant for diagnostic assistance ( P  = 0.007). SUMMARY: AI is closest to translation as diagnostic assistance during colposcopy. PosthrHPV triage - not primary screening - is the critical evidence gap. Future work should prioritize prospective multicentre multimodal (HPV + AI cytology + orthogonal biomarker) risk-calibrated triage models over further modality-pooled accuracy studies.