Hospital readmission rates for diabetes patients persist as a significant concern in healthcare, frequently signifying inadequate treatment results and elevated financial burden. This study introduces a complete ML framework that combines robust feature engineering with XAI to improve the prediction and interpretability of diabetes-related hospital readmissions. Extensive preprocessing, encoding, and class balancing were conducted on the Diabetes 130-US hospitals dataset to provide an enhanced feature space. Several machine learning models were assessed, with LightGBM attaining the greatest accuracy of 97.6%. The study utilized global and local explainable artificial intelligence tools, including SHAP, LIME, PDP, PFI, and DiCE, to mitigate the interpretability gap in machine learning predictions. These techniques yielded both aggregate and instance-specific insights into feature contributions, pinpointing critical predictors such as inpatient frequency, discharge disposition, and duration of hospital stay. The incorporation of XAI enhanced model transparency and enabled meaningful clinical interpretations. This study provides a reproducible and interpretable methodology for identifying early readmission risks, facilitating data-driven, patient-centered treatment in diabetes control.