Answers are grounded in this paper and cite their sources.
Unplanned 30-day hospital readmission remains a persistent, costly challenge in diabetes care, straining healthcare systems globally. Predictive models trained on clinical records can identify high-risk patients before discharge, but severe class imbalance between readmitted and non-readmitted cases causes standard classifiers to systematically overlook patients they should detect. This study develops and evaluates a five-phase hybrid pipeline for minority-class detection in 30-day diabetic readmission prediction, using the UCI Diabetes 130-US Hospitals dataset (101,766 inpatient encounters, approximately 8:1 class ratio). The framework sequentially applies Recursive Feature Elimination, BorderlineSMOTE, Conditional Tabular GAN augmentation, Edited Nearest Neighbours noise pruning, and CatBoost classification, benchmarked against three baselines – standard SMOTE, CTGAN alone, and Sequential SMOTE-CTGAN – trained without cost-sensitive tuning. The proposed framework achieved a minority-class recall of 60.15% and a Geometric Mean of 0.6279, versus 7.84%, 5.86%, and 7.31% recall across the baselines, none exceeding a G-Mean of 0.2778. Minority-class F1-Score improved from a best-baseline ceiling of 0.1306 to 0.2768, while overall accuracy declined from 88.68% to 64.94%, reflecting a deliberate tradeoff toward minority-class sensitivity over majority-class correctness. The full cost-sensitive pipeline recovered substantially more true readmissions than default-threshold oversampling baselines, although its ranking metrics were not superior, so broader validation is needed.
Loading...
Jharbade et al. (2026) studied this question.