EARLY DETECTION OF DIABETES USING ENSEMBLE MACHINE LEARNING AND EXPLAINABLE ARTIFICIAL INTELLIGENCE

Abstract

Diabetes mellitus is a chronic metabolic disorder which has been rapidly increasing and is a significant health burden in the world. Early detection is crucial to prevent serious complications and improve patient outcomes. Conventional diagnostic methods are based on clinical experience and lab investigations, which may not be efficient, scalable and interpretable for the risk assessment at early screening. The proposed framework uses PIMA Indians Diabetes Dataset which has 768 patients' records with 8 clinical variables and a binary outcome of diabetes. The data has been obtained from adult female patients of Pima Indian origin. Data preparation is done in a leakage free pipeline with invalid zero values considered as missing and substituted by median value, outliers are capped by the interquartile-range method, features are normalised by Min–Max and only the training set is balanced by the SMOTE. The accuracy, precision, recall, F1-score and ROC–AUC are used to evaluate the performance of ten predictive models: LR, DT, RF, SVM, KNN, GB, XGBoost, LightGBM, soft voting ensemble, and Deep Neural Network. The experimental results show that soft-voting ensemble and Random Forest are the most stable results with ROC–AUC about 0.81–0.82 . The Glucose, BMI and Age variables are found to be the most important variables in the model from the SHAP and LIME plots. The proposed system enhances the predictive accuracy and provides decision support for early detection of diabetes that is easy to interpret and clinically meaningful.