Forecasting 30-day hospital readmission is a major concern in healthcare because of its financial cost and the missed opportunity to act early on high-risk patients. We highlight the accuracy trap inherent in naive modeling of highly imbalanced data: a typical SMOTE-based Random Forest baseline reached a misleading 91% accuracy yet identified almost no high-risk patients (recall ≈ 0.0%). To address this, we built a comparative methodology that moves from cost-sensitive Random Forest classifiers to an optimized final model: XGBoost with class-balancing weights and post-hoc threshold tuning (lowered to 0.4396). The final model achieved a readmission recall of 69.97%, far above the baseline, offering a viable basis for clinical decision support in which identifying at-risk patients is the priority.