A leakage-aware and interpretable machine learning framework for ovarian cancer prediction using the PLCO screening dataset

Early diagnosis of ovarian cancer is a significant clinical problem, and this is mainly because there are no strongly sensitive early-detection biomarkers, and the standard screening parameters like CA-125 blood tests and transvaginal ultrasounds are very insensitive in making a diagnosis. This paper presents a machine learning architecture, which is organized and interpretable based on the Prostate, Lung, Colorectal, and Ovarian (PLCO) Cancer Screening data to enhance the risk stratification of ovarian cancer. The original analytical cohort comprised 78,209 participants, including 518 ovarian cancer cases. To select 20 informative pre-diagnostic predictors in a high dimensional feature space, a comprehensive feature selection pipeline with ANOVA (f classifier), Chi-square, Random Forest importance, and XGBoost importance were used. Several classification algorithms were tested and the boosting based methods proved to have a good predictive value. A stacking ensemble of K-Nearest Neighbors and AdaBoost was the highest performing overall (accuracy: 0.7644), and a test-set AUROC of 0.8399, with additional evaluation using sensitivity, specificity, PPV, NPV, balanced accuracy, MCC, and PR-AUC. SHAP analysis and permutation importance were used to provide model interpretability and consistently regarded questionnaire-based screening variables, age, and CA-125 attributes as the strongest ones. The suggested framework offers a methodologically rigorous and transparent method of assessment of population level ovarian cancer risk, providing a foundation for future research toward a clinically useful ovarian cancer screening decision-support system, pending external validation.