Screening Performance and Implementation Determinants of an AI-Assisted Visual Inspection application for Cervical Cancer Screening in Rwanda

Background: Cervical cancer remains a leading cause of preventable cancer death in low-and middle-income countries, where visual inspection with acetic acid (VIA) is widely used but highly dependent on provider expertise. Artificial intelligence (AI) could improve the consistency of VIA interpretation, yet evidence on real-world performance in African primary care settings remains limited. This study evaluated the screening performance of an AI-assisted VIA (AI-VIA) mobile application and identified implementation facilitators and barriers in Rwanda. Methods: We conducted a sequential mixed-methods study with an implementation science lens at Kicukiro Health Center, Kigali, Rwanda, from June 18 to July 25, 2025. Expert panel consensus served as the reference standard in the absence of histopathology. Quantitative analysis compared the screening performance of the AI algorithm against clinical expert assessment. Diagnostic accuracy measures such as sensitivity, specificity, positive predictive value and negative predictive value were analyzed, with inter-rater agreement assessed using Cohen's kappa. Qualitative data from a stakeholder focus group discussion were analyzed using the Consolidated Framework for Implementation Research (CFIR) to map implementation determinants. High-priority barriers and facilitators were identified and mapped to the Expert Recommendations for Implementing Change (ERIC) compilation to identify candidate strategies supporting broader implementation. Results: Of 251 women analyzed (mean age 35.6 years, SD 6.3), 65.7% had primary education and 88.1% were unemployed. Relative to expert consensus, nurse assessment demonstrated higher screening performance than the AI algorithm, with sensitivity 51.5% (95% CI 33.5–69.2) and specificity 98.2% (95% CI 95.4–99.5) versus sensitivity 33.3% (95% CI 18.0–51.8) and specificity 92.2% (95% CI 87.8–95.4). CFIR analysis identified facilitators including national policy alignment with cervical cancer elimination plan, interoperability, provider enthusiasm for objective decision support, and strategic positioning in the HPV-triage pathway. Critical barriers included algorithm sensitivity requiring dataset expansion, regulatory approvals, connectivity constraints, and concerns about human functions being replaced by AI. Conclusions: AI‑assisted VIA is feasible as HPV‑triage in Rwanda, given high specificity but low sensitivity. Scale‑up depends on model optimization with local, histology‑confirmed data, interoperability, clear regulation, and workforce engagement.