Diabetes is a common but serious health problem around the world that requires early and accurate prediction to avoid serious problems. Traditional diagnostics use a lot of resources hence machine learning can be a strong, data-driven way to make better predictions detecting diabetes. Using advanced data pre-processing to enhance data quality and predictability, the paper proposes a machine learning framework for diabetes prediction. The framework rigorously tests and optimizes various top-ranking classification algorithms, such as XGBoost, Random Forest, and an advanced ensemble learning technique, on Behavioral Risk Factor Surveillance System and Pima Indian Diabetes Dataset. Experimental results show that the designed ensemble model outperforms with higher performance, obtaining an accuracy of 83.82% and 88.28% and an F1-score of 84.34% and 86.00% for Datasets 1 and 2 respectively, overwhelmingly better than individual classifiers and competitive baselines. This framework has an extremely accurate and efficient aid for clinicians, eventually aiding timely interventions and enhancing patient outcomes in diabetes care.