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

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning

Background:

• Cervical cancer screening relies on detecting precancerous lesions (CIN2+).
• Accurate risk stratification is crucial for effective colposcopy triage.
• Existing models may lack explainability or external validation.

Purpose of the Study:

• Develop and validate a calibrated, explainable risk-stratification model for CIN2+.
• Utilize routine structured clinical data for broad applicability.
• Compare machine learning algorithms for optimal performance.

Main Methods:

• Retrospective study with development (879 women) and external validation (103 women) cohorts.
• Included 12 routine variables: demographics, HPV, cytology, colposcopy.
• Compared Logistic Regression, LightGBM, and XGBoost using cross-validation; XGBoost selected after calibration and threshold optimization.

Main Results:

• Calibrated XGBoost model achieved AUROC 0.720 (internal) and 0.679 (external).
• SHAP analysis identified cytology grade, HPV16, high-risk HPV, colposcopy impression, transformation zone, and age as key predictors.
• Model demonstrated clinical interpretability and supportive validation.

Conclusions:

• A calibrated machine learning workflow using routine data provides interpretable CIN2+ risk estimates.
• This model supports risk-based colposcopy triage, identifying women needing closer evaluation.
• The workflow offers a basis for prospective assessment and improved screening protocols.