A Benchmark and Meta-Analysis of Externally Validated AI Models for Cervical Cancer

This study aimed to establish a rigorous benchmark of externally validated artificial intelligence (AI) models for cervical cancer to assess their clinical readiness and generalizability across key clinical tasks. We conducted a systematic review and meta-analysis following PRISMA guidelines, searching PubMed/MEDLINE and Embase for studies developing or validating AI models in cervical cancer. Inclusion was restricted to studies reporting performance on held-out external test sets. Four clinical tasks were evaluated: screening/triage for CIN2+/CIN3+, preoperative staging surrogates (lymph node metastasis, parametrial invasion), prognosis (overall/disease-free survival), and treatment response prediction. Data on diagnostic accuracy, C-indices, calibration, and decision-curve analysis were extracted. Risk of bias was assessed using QUADAS-2 and PROBAST. From 1,662 identified records, 17 studies met the inclusion criteria. For screening/triage, four externally validated studies demonstrated that AI assistance significantly improved clinician sensitivity (Δ +17.85%) and specificity (Δ +4.52%), while multimodal fusion with HPV genotyping achieved AUCs up to 0.89. Four staging studies showed CT and MRI-based models with external AUCs around 0.86, incorporating calibration and decision-curve analysis. Only one externally validated prognostic model was identified, showing attenuated but maintained discrimination (C-index 0.66-0.70 externally). No externally validated treatment response models were found. The benchmark reveals a stratified landscape of AI readiness in cervical cancer. Externally validated evidence supports the clinical potential of AI for screening and staging, while prognostic and treatment response modeling require further multicenter validation. External validation and standardized reporting, beyond discrimination metrics alone, are critical for translating promising algorithms into reliable clinical tools.