Modern XGBoost-CNN Hybrid Stacking for Diabetes Risk Prediction Using BRFSS 2021

Diabetes, being an inveterate chronic condition, creates a serious worldwide health challenge. Therefore, its reliable and accurate forecast is essential for significant disease management and timely intervention. The publicly available 2021 BRFSS (Behavioral Risk Factor Surveillance System) dataset from Kaggle is used to mitigate the constraint of conventional models such as XGBoost, LightGBM, and Logistic Regression, which often fail to catch non-linear feature interactions and interdependencies among base learners. To address these difficulties, a strong stacking structure is put forward that combines modern XGBoost with histogram-based trees, LightGBM, and convolutional neural network (CNN)derived meta-features. To overcome data imbalance, multiple resampling methods were tested, including Random Oversampling, ADASYN, SMOTE, and the hybrid SMOTEENN approach, among which SMOTEENN exhibited outstanding performance in balancing class distribution while holding back data integrity. The outputs from base learners were subsequently combined using a Multi-Layer Perceptron (MLP), serving as a non-linear meta-learner, thereby increasing the model's generalization ability. Moreover, SHAP (SHapley Additive exPlanations) analysis was implemented to combine model behavior and identify the most significant features presenting to diabetes risk prediction. The outputs disclosed that features such as age, blood pressure, and body mass index (BMI) played a prominent role in classification results, improving the model's transparency and interpretability. The proposed hybrid structure attained exemplary metrics, including an accuracy of 0.951, an AUC of 0.907, a precision of 0.965, a recall of 0.983, and an F1-score of 0.974. These outputs displayed the advantages of integrating advanced gradient boosting algorithms, CNN-based meta-features, and non-linear ensemble integration for increased diabetes detection. In short, this research not only prominently enhances predictive performance but also provides a encouraging path for individualized healthcare guidance and early clinical decision support in diabetes management through the approach of modern machine learning and model fusing strategies.