Polycystic Ovary Syndrome (PCOS), a common hormonal condition affecting 6–12% of women in their reproductive years, poses significant diagnostic difficulties due to its diverse clinical presentations and overlapping symptoms. This study proposes a pioneering hybrid AI framework to enhance PCOS diagnosis, integrating four sophisticated fusion models: Fusion Model 1, a soft-voting ensemble of XGBoost, Random Forest, LightGBM, and CatBoost; Fusion Model 2, a convolutional neural network incorporating ResNet, VGG, and Inception-inspired architectures; Fusion Model 3, an attention-reinforced neural model emphasizing feature relevance; and Fusion Model 4, a voting ensemble combining K-Nearest Neighbors, Gaussian Naive Bayes, and Logistic Regression. To address data scarcity and class imbalance, we augment the dataset to 6,000 samples using Synthetic Minority Oversampling Technique (SMOTE), perturbation, interpolation, and Gaussian Mixture Models, ensuring robust representation of PCOS biomarkers. Employing 10-fold cross-validation, we rigorously assess model performance across nine metrics: AUC-ROC, Precision, Recall or Sensitivity, F1 Score, R2 Score, Cohen’s Kappa, Matthews Correlation Coefficient, Accuracy, and Balanced Accuracy. Our findings demonstrate superior predictive accuracy, with feature importance analysis underscoring the critical roles of testosterone levels, antral follicle counts, and menstrual irregularity.