Automated PCOS Disease Detection Using Clinical and Diagnostic Features

Abstract

Polycystic Ovary Syndrome (PCOS) is a prevalent endocrine disorder affecting reproductive-age women, leading to infertility, hormonal imbalance, insulin resistance, and cardiovascular complications. Early diagnosis remains challenging due to heterogeneous manifestations, overlapping symptoms, and lack of automated screening tools. To address these issues, this study presents a comprehensive comparative framework for PCOS prediction using machine learning and deep learning on a public dataset of 541 patient records. The framework incorporates missing value imputation, feature standardization, SMOTE class balancing, and correlation-based feature selection. Five machine learning algorithms (Decision Tree, KNN, SVM, Random Forest, XGBoost) and two deep learning architectures (LSTM, CNN-ResNet) were evaluated using accuracy, precision, recall, F1-score, and ROC-AUC, with confusion matrices employed for detailed classification analysis. KNN achieved the highest accuracy (92.66\%), recall (91.67\%), and F1-score (89.19\%), while XGBoost delivered the best discriminative capability (ROC-AUC = 0.9540). Among deep learning models, LSTM consistently outperformed CNN-ResNet, demonstrating superior ability to capture complex clinical feature relationships. SHAP-based explainability identified ovarian follicle counts, menstrual irregularities, hair growth, weight gain, skin darkening, and anti-Müllerian hormone (AMH) levels as the most influential predictors. These findings indicate that explainable machine learning models, particularly KNN and XGBoost, provide accurate and interpretable decision support for early PCOS screening, enabling timely intervention and offering a promising foundation for intelligent healthcare decision-support systems.