Cervical cancer remains the fourth most commonly diagnosed cancer in women globally, with an estimated 660,000 new cases and 350,000 deaths reported in 2022. Early risk stratification is constrained by severe class imbalance in available clinical datasets and by limited model interpretability. This study presents a systematic, reproducible framework comprising three integrated contributions. First, a head-to-head comparison of four class imbalance handling strategies Synthetic Minority Oversampling Technique (SMOTE), SMOTE-Tomek, Adaptive Synthetic Sampling (ADASYN), and class weighting is conducted across Random Forest (RF) and XGBoost classifiers using Repeated Stratified K Fold cross validation (K = 5, 10 repeats, 50 evaluations per model for the five model comparison; K = 5, 5 repeats, 25 evaluations per configuration for the imbalance strategy comparison) with Wilcoxon signed rank statistical testing. Second, five machine learning classifiers, Logistic Regression, Decision Tree, RF, Support Vector Machine (SVM), and XGBoost are evaluated on a held-out test set under the SMOTE pipeline across eight performance metrics. Third, SHAP TreeExplainer is integrated with a leakage safe out of fold (OOF) probability thresholding procedure to partition patients into clinically actionable Low, Medium, and High risk tiers. Experiments are conducted on the UCI Cervical Cancer (Risk Factors) dataset (n = 858; positive rate 6.4%; imbalance ratio 14.6:1). XGBoost achieved the highest test set AUC of 0.964, followed by RF (AUC=0.952). For RF, oversampling strategies (SMOTE, SMOTE-Tomek, ADASYN) improved mean cross-validation recall by up to 17.8 percentage points over no resampling (ADASYN recall: 0.731 vs baseline: 0.553). For XGBoost, class weighting produced the highest recall (0.706). At a clinically motivated classification threshold of 0.05, RF recall increased from 0.727 to 0.909 (TP: 10/11 cancer cases). Bootstrap confidence intervals (1000 resamples, 95% CI) confirmed result stability: RF AUC CI=0.868-0.999, XGBoost AUC CI=0.922 0.995. The uncalibrated RF model yielded a Brier score of 0.027, outperforming its Platt-calibrated variant (Brier =0.033) and the no-skill baseline (0.25). SHAP analysis identified the Hinselmann colposcopy result (mean |ϕ| = 0.084), Schiller test (0.071), number of pregnancies (0.048), and Citology result (0.046) as the dominant risk factors. The High risk tier (n = 15) exhibited 60.0% cancer prevalence, a 9.38 fold enrichment over the 6.4% population baseline. Robustness was confirmed via row-order holdout and age based subgroup generalisation experiments.
Sarkar et al. (Wed,) studied this question.