INTERPRETABLE ENSEMBLE FRAMEWORK FOR PCOS RISK PREDICTION WITH SHAP AND LIME-BASED EXPLAINABILITY

Abstract

Polycystic Ovary Syndrome (PCOS) affects an estimated one in ten women of reproductive age and presents with a heterogeneous range of metabolic, endocrine, and reproductive symptoms, complicating consistent clinical diagnosis. Methodology: Two related PCOS datasets (2,000 and 541 samples) were cleaned and combined for model training and external validation. Recursive Feature Elimination with Cross-Validation was used for feature selection, and class imbalance was addressed using the Synthetic Minority Over-sampling Technique (SMOTE), applied within each cross-validation fold to avoid data leakage. Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR), and a soft-voting ensemble of the three were trained and evaluated using accuracy, sensitivity, specificity, precision, F1-score, and ROC-AUC, with statistical significance assessed via McNemar's test. Results: On the external test set, SVM, RF, and the ensemble achieved comparable accuracy (96.7-97.2%), all significantly outperforming LR (90.5%; p < 0.0001), with no significant difference among the three (p > 0.05). SHAP and LIME were applied for global and local interpretability, identifying follicle count, weight gain, skin darkening, and hormonal markers (AMH, cycle regularity) as leading predictors, though follicle count's dominance also raises a diagnostic-circularity concern given its role as a clinical diagnostic criterion. Conclusion: This framework demonstrates that ensemble and tree-based models substantially outperform linear baselines for PCOS risk prediction, with the ensemble offering comparable predictive performance alongside interpretable, dual-level explanations. These findings contribute toward more transparent and clinically trustworthy machine learning tools for PCOS diagnosis support.