This study presents a machine learning-based framework for diabetes prediction using the Kaggle Pima Indian Diabetes dataset. The research integrates Random Forest, Synthetic Minority Oversampling Technique (SMOTE), and SHAP (SHapley Additive exPlanations) to improve predictive accuracy, interpretability, and class balance. After comprehensive data cleaning and feature engineering, including the creation of medically informed interaction variables, six supervised algorithms were compared. Random Forest demonstrated the best performance (AUC = 0.91), effectively capturing nonlinear relationships and providing clinically interpretable insights. SHAP analysis identified glucose, BMI, and family history as dominant predictors, aligning with known pathophysiological mechanisms. The proposed model’s interpretability and robustness make it suitable for clinical decision support and real-time mobile health applications. This study contributes to bridging the gap between algorithmic performance and clinical usability while highlighting the importance of balanced, transparent, and scalable approaches for early diabetes risk prediction and personalized healthcare management.