Class imbalance in diabetic readmission prediction: hierarchical SMOTE-GAN synthesis with stacking ensemble and dynamic cost-sensitive learning

Early prediction of 30-day hospital readmission in diabetic patients is clinically critical due to severe class imbalance (minority prevalence ≈9.2%). Existing oversampling approaches yield near-zero minority-class F1 scores via global synthesis, neglecting hard-to-classify boundary regions. We propose H-SMOTE-GAN+CS, comprising: (1) hierarchical two-level synthesis combining macro-level SMOTE with micro-level Wasserstein GAN with gradient penalty (WGAN-GP) trained on hard boundary samples, (2) a dynamic cost-sensitive stacking ensemble of XGBoost, Guided SGD ANN (GSGD-ANN), and Random Forest with false-negative penalty coupled to GAN quality metric D¯, and (3) Platt-scaled calibration with Youden-optimal threshold selection. Evaluated on the UCI Diabetes dataset (101,766 encounters), the pipeline achieves G−Mean=0.608 (+165%), minority-class F1=0.249 (+301%), Sensitivity=0.602 (+502%), AUC−ROC=0.656 (+7.5%), PR−AUC=0.188 (2.04× no-skill baseline), and Screening NNT=4.56 (2.83× improvement versus SMOTE-only methods), all significantly outperforming the GSGD-ANN+SMOTE baseline (p<0.001, 30 runs). The hierarchical synthesis stage alone achieves statistically equivalent discriminative performance (p>0.05, Wilcoxon signed-rank); the dynamic cost matrix contributes deployment configurability rather than discriminative improvement. Ablation confirms hierarchical synthesis as the dominant performance driver. SHAP attribution identifies inpatient visit count (0.443), time in hospital (0.171), and admission type (0.168) as leading predictors.