Improving diagnostic accuracy for PCOS: A hybrid machine learning architecture with feature selection, data balancing, and explainable AI techniques

Polycystic Ovary Syndrome (PCOS), which affects 5–10 % of women worldwide who are of reproductive age, is often misdiagnosed (∼70 %) despite the rising risks of metabolic disorders and infertility. Current machine learning diagnostics frequently struggle with unbalanced data and are not interpretable. This research improves PCOS diagnosis by introducing a new, interpretable hybrid architecture. We used mutual information and extra trees to improve feature selection and extensive preprocessing, including SMOTE for class imbalance, on a dataset of 541 patient records. A Soft Voting Ensemble that included Multilayer Perceptron (MLP) with CatBoost, optimized using GridSearchCV, and verified with 5-fold cross-validation, outperformed each individual model with previous research with a state-of-the-art accuracy of 96.88 %. Additionally, deep learning models performed well, most notably DANet (94.50 % accuracy). Importantly, SHAP and LIME improved model interpretability, offering clear insights into diagnostic judgments. The architecture was put into practice in an intuitive Flask web application for explainable, real-time forecasts. This study offers a therapeutically applicable method that strikes a balance between interpretability and high accuracy, enabling early PCOS identification and better patient outcomes. Multimodal integration and dataset extension are potential avenues for future study.