Diabetes Prediction Using Optimized Ensemble Machine Learning on Clinical and Lifestyle Indicators

Abstract

Early diagnosis of diabetes is important in reducing the associated health risks and in enabling quick medical interventions. This paper presents a machine learning forecasting model based on an interpretable framework using the Diabetes Health Indicators Dataset on Kaggle. The data set contains clinical and lifestyle health variables that act as input attributes of prediction. The classification of persons at risk was done using different algorithms including Logistic Regression, Random Forest, Decision Tree, CatBoost, and an ensemble Voting Classifier that combines XGBoost and LightGBM. To overcome the issue of data imbalance in the advanced models, Synthetic Minority Oversampling Technique (SMOTE) was applied to increase class representation and model generalization. Explainable Artificial Intelligence (XAI) interpreters, LIME and SHAP were added to provide both local and global interpretability, explaining the most important health variables influencing the prediction of diabetes. The Voting Classifier achieved the best accuracy of 92 with respect to single models such as Logistic Regression and the Random Forest, thus demonstrating the effectiveness of combining sophisticated resampling and ensemble techniques in making transparent and high-performance predictive models in healthcare analytics