<![CDATA[This study examines the effects of SMOTE and SMOTE-TomekLink on XGBoost classification and SHAP interpretability using an imbalanced diabetes dataset (100,000 samples, 91.5% : 8.5% ratio). Three research gaps are addressed: (i) direct experimental comparison of both resampling techniques within identical pipelines; (ii) quantification of minority sensitivity stability through Stratified 5-Fold Cross-Validation; and (iii) reliability of post-resampling SHAP interpretation. Results demonstrate significant superiority of SMOTE-TomekLink with a F1-score of 0.89 (SMOTE 0.85, baseline 0.81), recall of 0.86 (SMOTE 0.85, baseline 0.70), and the lowest false positive rate of 176 cases (SMOTE 345, baseline 40). The lowest cross-validation standard deviation (0.0021) validates superior stability. These figures meet the very good performance category (F1-score >0.85, recall >0.80) for imbalanced medical classification. Primary contribution: the first controlled comparative design that simultaneously examines resampling effects on classification metrics and SHAP interpretation validity. SHAP analysis verifies HbA1c and blood glucose as dominant predictors (52.3% variance), consistent with medical standards.]]>
Copyrights © 2026
