A consensus-based interpretable machine learning framework for robust early diagnosis of ovarian cancer using hematological and tumor biomarkers

The early and accurate diagnosis of any type of ovarian cancer is one of the key clinical issues because of the heterogeneous expression of biomarkers and the high prices of false-negative diagnosis. The proposed study is a consensus-based machine learning framework that can improve diagnostic stability, interpretability and clinical utility by utilizing routine hematological and tumor marker data. They experimentally tested 15 machine learning classifiers, under three feature selection paradigms: Stability Selection, BorutaSHAP, and a new Consensus Feature Selection paradigm that combines filter and wrapper feature selection. Stratified five-fold cross-validation and independent test set were used to determine its model performance and the measurement scales such as accuracy, sensitivity, specificity, ROC-AUC, calibration, and analysis of decision curve. A Stochastic Gradient Descent (SGD) Classifier using Consensus Feature Selection gave the strongest performance, with 97.14% accuracy (95% CI 90.17–99.21%), 100.00% sensitivity (95% CI 90.36–100%), and 94.12% specificity (95% CI 80.91–98.37%), with no false-negative examples among all configurations. Clinical reliability was also established on the basis of probability calibration (Brier score = 0.088), decision curve analysis with better net benefit at clinically relevant thresholds, and robustness to Gaussian noise perturbations up to 20%. SHapley Additive exPlanations (SHAP) were employed to guarantee model interpretability with the most significant role of HE4, CA125 and menopausal status as shown in the existing clinical knowledge. The proposed pipeline, while constrained by sample size and single-center data, demonstrates strong proof-of-concept performance and may inform future clinical decision support and precision oncology processes, pending prospective multi-center validation.