Early-Stage Diabetes Mellitus Risk Prediction Using Light Gradient-Boosting Machine With Synthetic Minority Oversampling Technique Combined With Edited Nearest Neighbors on a Novel Southern Indian Clinical Dataset

India has the world's second-highest diabetes rate. According to the most recent survey, 25 million people are at risk of getting type 2 diabetes, while 77 million are already suffering from it. Diabetes is one of the leading chronic diseases in India and around the world. Early diagnosis is essential for effective disease management and prevention of complications. However, current diagnostic methods fail to detect early-stage risk factors. This research introduces a region-specific, clinically validated dataset and proposes a Light Gradient-Boosting Machine (LightGBM)-based framework with Synthetic Minority Oversampling Technique combined with Edited Nearest Neighbors (SMOTE-ENN) for early-stage diabetes risk prediction. We proposed a novel Southern India Diabetes Dataset by curating clinical records from the Koppal Institute of Medical Sciences in Koppal and Dhanavantari Multispeciality Hospital, Jamkhandi, Karnataka, India, encompassing 1,185 patient data annotated with 17 clinical and demographic characteristics. To balance the dataset's intrinsic class imbalance, the SMOTE-ENN approach was employed to improve model generalization. Six models, such as Random Forest, Gradient Boosting, XGBoost, LightGBM, CatBoost, and Multilayer Perceptron, were evaluated, followed by a comparative performance analysis. Among these, LightGBM gives better performance, achieving a train accuracy of 97.09%, test accuracy of 97%, and area under the receiver operating characteristic curve of 0.99. These experimental results emphasize the potential of integrating hybrid resembling techniques with gradient boosting frameworks for accurate and clinically relevant early-stage diabetes risk prediction.