Ensemble-Based Machine Learning Framework for Early Diabetes Prediction Using Localized and Normalized Data

Abstract

Diabetes is a serious condition that requires early diagnosis to minimize the long-term effects, but the traditional methods of diagnosis usually diagnose it at a late stage. Machine learning has potential, but generic models often are not able to accommodate population-specific health variations. We present a machine learning model in the form of an ensemble that uses (i) normalization to standardize heterogeneous medical data, (ii) location-specific corrections and correlation-based feature reweighting to focus on clinically relevant predictors and (iii) a combination of Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost) using a soft-voting system. It was tested on the PIMA Indians Diabetes Dataset, where the performance was measured in terms of confusion matrices, ROC curves, and standard measures The ensemble structure proved to be better at predicting than baseline models (Logistic Regression, Naive Bayes, and KNN). XGBoost was most accurate and had the highest AUC (89.1 and 0.93) with individual classifiers, whereas the ensemble misclassified fewer individuals and had better recall among minority (diabetic) cases. Through preprocessing with location sensitivity and feature weighting with correlation, the proposed framework is successful in capturing subtle risk patterns and implicitly solving the problem of class imbalance. It has high stability and can be generalized, making it appropriate to implement in the real healthcare setting The ensemble-based model proposed provides a viable and scalable model to predict diabetes at an early age. The next step in the work will be the extension of the model to real-time patient data and its implementation in clinical and mobile health platforms to become more widely accessible.