Calibrated and Explainable CIN2+ Risk Stratification Using Routine Clinical Data: Development and External Validation

To develop and externally validate a calibrated, explainable CIN2+ risk-stratification model using routine structured clinical data.

This retrospective study included 879 women in a development cohort and 103 women in an independent external validation cohort. Twelve routinely available variables covering demographic status, reproductive history, HPV status, cytology, and colposcopic findings were used. Logistic regression, LightGBM, and XGBoost were compared using stratified 5-fold cross-validation with pooled out-of-fold predictions. All preprocessing, calibration, and threshold-selection steps were fitted within training folds to reduce information leakage. Performance was assessed by discrimination, probability calibration, decision curve analysis, threshold-based operating characteristics, and SHAP-based interpretability.

XGBoost provided the most favorable overall balance and was selected as the primary model after isotonic calibration. In internal validation, the calibrated model achieved an AUROC of 0.720, an AUPRC of 0.496, and a Brier score of 0.172. A high-sensitivity threshold favored detection of CIN2+, whereas a balanced threshold provided a more even trade-off between sensitivity and specificity. In external validation, the model achieved an AUROC of 0.679, an AUPRC of 0.505, and a Brier score of 0.202. SHAP identified cytology grade, HPV16, overall high-risk HPV positivity, colposcopic impression, transformation zone type, and age as leading contributors.

A calibrated machine-learning workflow based on routine clinical variables yielded clinically interpretable CIN2+ risk estimates with supportive internal and external validation. By combining calibration, threshold-based reporting, and decision utility, this workflow may support risk-based colposcopy triage by helping identify women who require closer evaluation while providing a basis for subsequent prospective assessment.