Diabetes, a chronic metabolic disease, is caused by persistently high blood sugar levels and requires early detection for effective management. However, the limited amount of labeled data and the presence of outliers (missing values) in diabetes datasets make reliable predictions challenging. In this work, we propose a robust framework, HDP (Human Diabetes Prediction), for diabetes prediction based on tuned ensemble models, particularly the gradient boosting classifier and tuned gradient boosting classifier, which incorporate hyperparameter optimization to improve prediction accuracy. We compare the performance of the HDP model with various existing machine learning models on four datasets, including the UCI Machine Learning PIMA Indians dataset, Cleaned PIMA Indians dataset, and Modified PIMA Indians dataset. The experimental results show that the proposed HDP model performs better than other models like Decision Trees, Logistic Regression, XGBoost, Support Vector Machine, Random Forest, and K-Nearest Neighbors.. The HDP model achieves 88%-90% accuracy across three datasets, offering a 2%-4% performance improvement over other models. This study demonstrates the effectiveness of the HDP model, highlighting its potential for improving diabetes prediction accuracy in healthcare applications.