Diabetes mellitus is a chronic metabolic disorder that poses a significant and escalating threat to global public health if not diagnosed and managed early. The asymptomatic nature of the disease in its preliminary stages often leads to delayed clinical diagnoses, resulting in severe long-term compli-cations. This study presents a highly optimized machine learning framework designed for the early and accurate detection of diabetes using the widely recognized PIMA Indians Diabetes dataset. Building upon foundational benchmarks established in prior literature, particularly the work of Jain et al., we propose a comprehensive two-phase methodology. Phase 1 rigorously replicates baseline performance across traditional Decision Tree, Naive Bayes, and Random Forest models to establish a work-ing control. Phase 2 introduces advanced data preprocessing techniques—specifically targeted median imputation for biolog-ically impossible missing values, exploratory data analysis, and hyperparameter tuning—alongside modern ensemble techniques including XGBoost, Gradient Boosting, and a Hybrid Soft-Voting Classifier. Our experimental results demonstrate that while baseline models achieve high raw accuracy, the integration of robust preprocessing combined with sophisticated ensemble methods significantly improves model reliability, generalizability, and diagnostic separability. The proposed optimized framework achieves a peak test accuracy of 83.12% and an exceptional ROC-AUC score of 0.9098, indicating near-perfect class separability and offering a highly reliable, computationally efficient tool for clinical diagnostic assistance.

Index Terms—Diabetes Prediction, Random Forest, Naive Bayes, XGBoost, Voting Classifier, Machine Learning, Predictive Analytics, PIMA Dataset.