Volume 2 - Issue 3, May - June 2026

Predicting diabetes accurately can be tough because the data used in medical studies often has an uneven distribution of cases, and the features in the data don't always follow a straight line pattern. Models like Random Forest and Logistic Regression can give reasonable results, but they often fail to detect the less common cases in datasets like the Pima Indians Diabetes Database. This research introduces a new approach that uses the Synthetic Minority Oversampling Technique to fix the imbalance in the data and Kernel Principal Component Analysis to better understand complex feature relationships. The method was tested with several machine learning techniques, including CatBoost, XGBoost, and Multi Layer Perceptrons. The results show that the CatBoost model, when combined with the suggested data processing steps, reached an accuracy of 82% and had strong ROC AUC scores. This is better than the results seen with traditional methods like Random Forest in earlier studies. The study shows that creating balanced data and handling complex features are key to building better and more accurate systems for predicting diabetes. It also emphasizes the need to address both class imbalance and complex data patterns in medical datasets to improve prediction outcomes.

Vishvas Gupta, Vinay Rawat, Tanish Agrawal, Vipul Sharma, Kashish Shrotriya, Varnika Datey, Tanisha Gupta, Rashmi Pandey,"An Optimized Predictive Framework for Diabetes Mellitus Using SMOTE, Kernel-PCA, and Stacking Ensemble Learning" International Journal of Advanced Multidisciplinary Research and Educational Development, V2(3): Page(682-687) May-June 2026. ISSN: 3107-6513. www.ijamred.com. Published by Scientific and Academic Research Publishing.