Polycystic Ovary Syndrome (PCOS) is recognised as a wide-spread and impactful condition affecting women’s health on a global scale. This study explores the application of three distinct deep learning architectures Deep Neural Network (DNN), Deep Cross and Network (DCN) and TabNet for early PCOS detection, leveraging a comprehensive data preprocessing pipeline. Key preprocessing steps included feature selection through Chi-square and Mutual Information scores, outlier removal via the Interquartile Range (IQR) method, label encoding, and feature transformation using the Yeo-Johnson power method to address skewed distributions and using SMOTE for handling class imbalance, ensuring fair model training. TabNet outperformed other models with balanced precision, recall, and F1-score. Explainable AI techniques such as SHAP and LIME were applied as they provide transparent insights into the model’s decision-making process. Our proposed study determines the effectiveness of combining advanced preprocessing, deep learning models, and explainable AI techniques in developing reliable and interpretable models for PCOS classification.