The global importance of Diabetes Mellitus (DM) is addressed by using advanced machine learning approaches to increase early detection and predictive precision. Due to the increasing prevalence of diabetes, especially in developing countries, the study shows that model applicability, algorithm choice, and long-term prediction accuracy are little understood. Dataset preprocessing—cleaning, scaling, and controlling outliers using the Interquartile Range (IQR) approach—is essential. Comparing XGBoost vs LightGBM shows differences in accuracy, precision, recall, F1 score, and AUC score. XGBoost with IQR preprocessing produce good accuracy, precision, and recall, making it a potential predictor. However, LightGBM has various performance indicators, highlighting the importance of considering environmental and application requirements when picking a model. To construct reliable diabetes prediction models, rigorous preprocessing and model validation are essential.