Abstract / Summary
Abstract Background: Cervical cancer is a leading malignancy and mortality cause among Ethiopian women. Despite screening expansions coverage remains low. Traditional models miss complex determinants. This study aimed to develop machine learning models to predict cervical cancer screening utilization and identify drivers using the 2024 to 2025 Ethiopian Demographic and Health Survey. Methods: A secondary analysis of the 2024 to 2025 Ethiopian Demographic and Health Survey was conducted among women aged 15 to 49. Six algorithms Logistic Regression, Decision Tree, Support Vector Machine, Random Forest, Gradient Boosting Machine and Extreme Gradient Boosting were trained. Synthetic Minority Over-sampling Technique handled class imbalance with a 4.2 percent screening prevalence. Models were evaluated on a holdout set using Area Under the Receiver Operating Characteristic curve, Area Under the Precision-Recall Curve, Brier score and Decision Curve Analysis. SHapley Additive exPlanations explained model predictions. Results: Extreme Gradient Boosting achieved superior discrimination with an Area Under the Receiver Operating Characteristic curve of 0.896 and Area Under the Precision-Recall Curve of 0.781 outperforming baseline Logistic Regression at 0.772. Isotonic calibration optimized prediction loss with a Brier score of 0.038. Decision Curve Analysis confirmed clinical net benefit across thresholds from 5 percent to 55 percent. Explanatory analysis identified cancer awareness, education, autonomy, wealth quintile, recent health visits and urban residence as primary screening drivers. Conclusions: Probability-calibrated Extreme Gradient Boosting with explainability provides an accurate tool for identifying un-screened women in Ethiopia. Public health programs can deploy these tools in primary care applications to target outreach, optimize community engagement and accelerate World Health Organization elimination targets.