A Hybrid Machine Learning Framework for Diabetes Prediction with Enhanced False Negative Reduction and Clinical Reliability

Abstract

Accurate and clinically reliable prediction of diabetes is essential for early diagnosis and prevention of severe complications. This paper proposes a hybrid machine learning framework for diabetes prediction with a specific focus on minimizing false negative errors, which are critical in clinical decision-making. The framework is evaluated on a large-scale dataset comprising 100,000 patient records with demographic, clinical, and biochemical features. A diverse set of models was employed, including Naive Bayes, Logistic Regression, K-Nearest Neighbors, Gradient Boosting, Support Vector Machine, Decision Tree, Random Forest, and advanced boosting algorithms such as LightGBM,  XGBoost, and CatBoost, along with a Tabular Neural Network. Three hybrid models were developed to enhance predictive performance. Among them, the neural network and XGBoost hybrid achieved the best results with an accuracy 98.45%, PR-AUC of 0.883 and ROC-AUC 97.82%. Feature analysis identified HbA1c and blood glucose level as dominant predictors. Probability calibration and threshold tuning improved sensitivity to 90.29%, significantly reducing false negatives. Detailed error analysis revealed that misclassified cases exhibit borderline clinical values and overlapping feature distributions. The proposed framework demonstrates improved robustness and suitability for real-world clinical decision support systems.