Robust and interpretable prediction of diabetes is still a challenging problem in healthcare analytics. In this paper, we propose a comprehensive machine learning framework for the diabetes prediction based on PIMA Indian Diabetes dataset. Five new models are introduced, each of which is developed to overcome some limitations in conventional classifications like non-transparency, inability in adaptation and unawareness of data uncertainty. The pipeline starts with a SHAP-aware Random Forest to set an interpretable baseline. It is further enhanced by contrastive feature enhancement for more informative representation, SHAP-guided reinforcement learning to select the better features, a physiological age-based model to achieve clinical alignment and an adversarial robustness analysis with synthetic patient profiles. All methods are compared in terms of interpretability, prediction accuracy and robustness to perturbation. The inclusion of explainable AI, domain-specific features and robustness testing provides a trustworthy diagnostic tool that can be rolled out. The proposed approach paves the way for future work in personalized medicine, providing a route towards intelligent, interpretable and clinically relevant decision-support systems.