Inam Abousaber ¹ Haitham F. Abdallah ^2* Hany El-Ghaish ³

• ¹ Dept. of Information Technology, Faculty of Computers and Information Technology, University of Tabuk, tabuk, Saudi Arabia
• ² Dept. of Electronics and Electrical Comm., Higher Institute of Engineering and Technology, Kafrelsheikh, Egypt, Kafr El Sheikh, Egypt
• ³ Dept. of Computer and Automatic Control, Faculty of Engineering, Tanta University, Tanta, Egypt

Diabetes prediction from clinical datasets is a critical challenge in medical data analysis, particularly due to the frequent imbalance in class distributions, where non-diabetic cases often dominate. Such imbalances can compromise the predictive performance of machine learning models, leading to biased outcomes and reduced generalization. To address these challenges, we present a novel predictive framework that leverages cuttingedge machine learning algorithms and sophisticated imbalance handling techniques. Our approach systematically evaluates various machine learning models, applying advanced feature engineering and resampling strategies to enhance predictive accuracy. We conduct rigorous testing on three widely recognized datasets-PIMA, Diabetes Dataset 2019, and BIT 2019 -demonstrating the robustness and adaptability of our methodology across different data environments. This work introduces a comprehensive experimental setup that underscores the importance of model selection and imbalance mitigation in achieving reliable and generalizable diabetes prediction outcomes. The findings contribute significant insights to the field of medical informatics, offering a robust, data-driven framework that improves prediction accuracy in the presence of challenging class imbalances, thereby advancing the state-of-the-art in diabetes prediction.