An explainable ensemble learning framework for ovarian cancer classification using blood biomarkers

Background Ovarian cancer remains one of the most lethal gynecological malignancies, largely due to delayed diagnosis and the limited sensitivity of conventional screening approaches. Methods This study proposes an interpretable machine learning framework for the binary classification of ovarian cancer and benign ovarian tumors using routine blood biomarker data. A soft voting ensemble combining Histogram-Based Gradient Boosting and K-Nearest Neighbors was developed to capture complementary global and local data patterns. Class imbalance was mitigated using the ROSE resampling technique, while LASSO-based embedded feature selection improved model robustness and generalization. Model transparency was ensured through a comprehensive explainable AI pipeline incorporating SHAP, LIME, anchor rules, ELI5, Partial Dependence Plots, and surrogate decision trees. Findings The proposed ensemble consistently outperformed eleven benchmark classifiers across multiple resampling strategies, achieving a maximum accuracy and F1-score of 98.61Interpretation Integrating ensemble learning with multi-level explainability yields a highly accurate, reliable, and clinically interpretable diagnostic tool, supporting its potential adoption in real-world ovarian cancer screening and decision-support systems.