Enhancing Diabetes Diagnosis with Confidence-Calibrated Adaptive Weighting and Multi-Model Ensemble Framework

Abstract

Metabolic conditions like diabetes often show up as high blood sugar levels, or hyperglycemia, which happens when the body doesn't make enough insulin. If this goes unnoticed or untreated, it can cause serious damage to critical organs such as the eyes, kidneys, nerves, heart, and blood vessels. Recently, machine learning and other computational techniques have shown great potential for predicting who might develop diabetes. But there's still room to improve how accurate and reliable these models are. In this work introduce a step-by-step machine learning approach to predict and diagnose diabetes using three different datasets. First, used Hot Deck Imputation (HDeckImp) to fill in missing data and make the classifications more accurate. Next, we apply K-fold cross-validation to test how well the model holds up with different data splits, ensuring robustness. Then, we incorporate multiple classifiers like Random Forest, XGBoost, AdaBoost, and Bagging to strengthen predictions. Finally, we developed a new weighting method called Confidence-Calibrated Adaptive Weighting (CCAW). This method dynamically gives more weight to models that are both accurate and confident, measured through information entropy, so the best models have more influence on the final outcome. Experiments show that this approach, especially with CCAW, reaches a top accuracy of 98.9% on the Frankfurt dataset, beating previous ensemble methods such as ADR-W. This improvement not only enhances prediction accuracy but also makes the model stable, which is essential for practical clinical use.