International Journal of Science, Engineering and TechnologyAuthors: Pooja K
Abstract: Cervical cancer remains one of the leading causes of cancer-related death among women, particularly in regions where organised screening is scarce. Machine learning can support early risk assessment, but many published models report accuracy alone, act as opaque “black boxes”, and are rarely delivered as usable tools. This study presents an explainable machine learning framework and web application for cervical cancer risk prediction that combines Recursive Feature Elimination (RFE) with the Extreme Gradient Boosting (XGBoost) algorithm. The pipeline uses the publicly available Cervical Cancer Risk Factors dataset (858 patients, 36 attributes, 6.4% biopsy-positive), median imputation, two engineered features (Lifetime Exposure Index and Cancer Load Score), a frequency-based feature filter, Synthetic Minority Over-sampling (SMOTE) applied to the training partition only, and RFE with a Random Forest estimator to retain 18 predictors. On a stratified hold-out set of 172 patients the model achieved an accuracy of 96.51%, specificity of 98.76%, recall of 0.636, precision of 0.778, ROC-AUC of 0.959 and a Matthews correlation coefficient of 0.686. Because the positive class is small, a leakage-free repeated stratified cross-validation (5 folds × 3 repeats, with SMOTE and RFE nested inside each fold) was also performed and is reported alongside Logistic Regression, Random Forest and XGBoost without RFE. SHAP analysis shows that the Schiller and Hinselmann colposcopic test results dominate the prediction, followed by reproductive history and the engineered exposure index. The model is deployed in a Streamlit application that returns a probability-based Low, Medium or High risk category, visual explanations, an analytics dashboard and a downloadable PDF report. The system is intended for decision support and awareness, not for diagnosis.
