Comprehensive guidelines for establishing sample diversity and data sufficiency in artificial intelligence diagnostic datasets for cervical liquid-based cytology

This is the authors' abstract. We don't add key points for this paper.

Liquid-based cytology (LBC) has become a core technology in cervical cancer screening, and artificial intelligence (AI) shows great potential in addressing issues such as the global shortage of cytopathologists and large variations in diagnostic results. However, the clinical reliability of AI systems in the field of cervical cytology fundamentally depends on the quality, diversity, and representativeness of their training and validation datasets. Drawing on evidence-based medicine principles, the latest research progress in China and globally, and clinical practices, these guidelines establish standardized requirements for sample diversity and data sufficiency in cervical LBC AI datasets to reduce algorithmic bias. The objective is to improve the real-world applicability of models and ensure their safe clinical application. This guideline strictly complies with standardized guideline development specifications, with transparent expert panel organization, systematic literature retrieval, standardized evidence grading, three-round Delphi expert consensus, external peer review, and dynamic update mechanisms. All quantitative threshold indicators in the recommendations are jointly formulated based on high-quality clinical evidence and expert consensus, with clear evidence sources and consensus construction processes to enhance the transparency, credibility, and operability of the guideline.