Diabetes mellitus is a rising global health concern, necessitating accurate and timely diagnosis to prevent serious complications. While machine learning (ML) has demonstrated considerable potential in diabetes classification, existing models often struggle with class imbalance, limited clinical interpretability, and insufficient benchmarking against advanced methods. To address these gaps, this study introduces a novel hybrid framework that integrates cost-sensitive learning, ADASYN-based adaptive sampling, and stacked ensemble modeling to enhance both classification accuracy and clinical relevance. The framework was evaluated using the PIMA Indian Diabetes dataset, under both balanced and imbalanced conditions, and benchmarked against state-of-theart models. The proposed ensemble, featuring cost-sensitive XGBoost combined with ADASYN, achieved 89.47 % accuracy and 82.72 % recall, outperforming traditional models. SHAP (SHapley Additive exPlanations) analysis was employed to ensure transparency, confirming that the model's key predictors, glucose, BMI, and age, are consistent with established medical knowledge. This work contributes to the field by offering methodological innovations that bridge the gap between machine learning (ML) research and real-world healthcare applications, thereby enhancing both performance and clinical interpretability through explainable AI. The study's findings highlight the potential of this approach for early, scalable, and clinically applicable diagnosis of diabetes.