Clinically benchmarked and interpretable machine learning for preoperative discrimination of malignant and benign ovarian tumors

Ovarian cancer is a highly fatal gynecological malignancy, largely because it is frequently diagnosed at an advanced stage. Although machine learning (ML) has shown promise for ovarian cancer classification using clinical laboratory data, many existing models operate as black boxes, limiting clinical interpretability. This study proposes an interpretability-centered ML framework for distinguishing malignant from benign ovarian tumors using a publicly available, retrospective, single-center Chinese cohort from the Third Affiliated Hospital of Soochow University, comprising 349 patients with 171 malignant and 178 benign ovarian tumors. Variables with excessive missingness were excluded, while retained missing values were mean-imputed and features were min–max normalized. Imputation statistics and normalization parameters were estimated exclusively from the training data and applied unchanged to the evaluation data to prevent information leakage. Five classifiers, support vector machine (SVM), k-nearest neighbor, discriminant analysis, decision tree, and logistic regression, were evaluated using Bayesian hyperparameter optimization. SVM achieved the highest accuracy of 93.7% and sensitivity of 95.3% considering the leave-one-out cross-validation. The proposed framework optimizes both predictive and explanation stages: Bayesian optimization first tunes the classifiers, followed by joint optimization of LIME perturbation sample size and kernel width. The optimized LIME surrogate achieved 91.9% accuracy, only 1.8% points below the SVM, while identifying CA125 and CA19-9 as the dominant positive predictors of malignancy, consistent with established clinical knowledge. These findings indicate that systematic calibration can improve the fidelity and stability of patient-level explanations, although stronger predictive results have been reported on the same dataset and external validation remains necessary. However, prospective, multicenter, and clinician-in-the-loop validation is required before clinical decision-support use.