An explainable machine learning approach for cervical cancer screening: decoding morphological diagnostic drivers.

Abstract

Background: Cervical cancer screening in primary care is hindered by expert pathologist shortages and heavy diagnostic workloads, leading to fatigue-induced misdiagnoses. This study evaluated the diagnostic capacity, subpopulation robustness, and operational efficiency of an interpretable machine learning (ML) tool within a large Chinese healthcare network.

Methods: A retrospective database of 5,000 women was audited. Archived liquid-based cytology (LBC) digital slides were evaluated via a parallel validation channel: the LBC-NET pipeline versus a double-blind panel of senior cytopathologists. Diagnostic safety was tested by stratifying data across age, vaginal infections, and menopause conditions. Algorithmic transparency was quantified using game-theoretic Shapley Additive exPlanations (SHAP) and morphological feature tracking. Operational efficiency was evaluated via continuous cumulative distribution curves.

Results: For detecting high-grade squamous intraepithelial lesions or worse (HSIL+), the ML model achieved a clinical sensitivity of 92.5%, significantly outperforming the manual method at 84.2% (p < 0.001), with an area under the curve (AUC) of 0.94. Subgroup stratification demonstrated that menopausal tissue atrophy or active vaginal infections caused no significant performance degradation (p > 0.05). SHAP analytics revealed that the nucleus-to-cytoplasmic ratio and nuclear hyperchromasia intensity were the core predictive metrics. Operationally, the ML system compressed the median slide-review time by 63.0%, escalating the daily screening capacity by 175% and reducing specialist referral consultations by 45.2%.

Conclusion: The interpretable ML tool provides a reliable and fast clinical choice for massive cervical cancer screening.