Clinical Weighting and Symptom Interaction-based Risk-Stratified Learning for Early Prediction of Ovarian Cancer Using Explainable Machine Learning

Objectives: To develop an explainable Clinical Weighting and Symptom Interaction-based Risk-Stratified Learning (CWSI-RSL) framework for early ovarian cancer prediction using clinical feature weighting, symptom interactions, and patient risk-level classification. Method: 349 records, 51 attributes ovarian cancer dataset was used. Missing values were treated and duplicate records were handled. Numerical data were normalised. Mutual Information (MI) and Recursive Feature Elimination (RFE) were used for important features selection. Clinical weights and Symptom Interaction Feature Expansion (SIFE) features were then added, and low-, medium- and high-risk grouping was performed. Logistic Regression (LR), Support Vector Machine (SVM), Naïve Bayes (NB) and K-Nearest Neighbours (KNN) were optimised using GridSearchCV by 80:20 train-test splitting. The performance achieved was compared against previously reported traditional machine-learning and ensemble baselines using accuracy, precision, recall and F1-score, and SHAP (SHapley Additive exPlanations) was employed to interpret the model predictions. Findings: The accuracy, precision, recall, and F1-score of CWSI-RSL were 92.40%, 90.30%, 95.00% and 92.60%, respectively. It performed 5.26 percentage points better than the best reported conventional model, Naïve Bayes (87.14%) and 5.40 percentage points better than the best reported ensemble model, Boosting (87.00%). The recall of 95.00% shows that the model was able to detect most of the ovarian cancer cases, and the F1-score of 92.60% shows a balanced performance. Menopause status, CA125-related interactions, and neutrophil-based interactions were identified as important predictors by SHAP analysis. These results demonstrate that combining clinical weighting, symptom interactions and risk grouping can improve prediction and provide useful explanations. Novelty: CWSI-RSL integrates clinical weighting, symptom interaction, model-derived risk grouping, optimised machine learning and SHAP explainability in a unified ovarian cancer prediction framework.