Diabetes is still a major health problem worldwide and efficient predictive modeling is necessary for early diagnosis and intervention. This paper assesses the performance of several machine learning models in predicting diabetes based on three different dataset. There are three datasets: PIMA Dataset, Indian Healthcare Diabetes Dataset and a Merged dataset combining both. A variety of supervised machine learning algorithms are implemented to test their ability to predict: Support Vector Machines (SVM), Decision Trees (DT), Random Forest (RF), XGBoost and Ensemble Methods. Comparing model performances show datasets variable variations to great magnitudes. Here on the Indian Healthcare dataset a 99.64% accuracy is achieved by the Random Forest classifier and Bagging classifier while 98.92% is given by Decision Tree Classifier Results further illustrate that a larger, merged dataset improves model generalization and increases predictive accuracy the importance of appropriate dataset selection in data collection for ML-based medical diagnostic The research suggests that the application of complicated ML techniques holds great promise in improving timely detection and early diagnosis of diabetes thus facilitating healthcare professionals to make more informed decisions.