 |
International Society of Science and Applied Technologies |
| Efficient Hospital Readmission Prediction with Reduced Feature Sets Under Class Imbalance | ||||
| Author |
Kehan Gao
|
|||
| Co-Author(s) |
Fatma Pakdil; Steve Muchiri; Garrett Dancik; Ece Pakdil; H. Nail Aydin
|
|||
| Abstract | Hospital readmissions remain a major challenge for
healthcare systems due to their impact on patient outcomes and
healthcare costs. Accurate prediction of high-risk patients can
support targeted interventions and reduce avoidable readmissions.
This study investigates hospital readmission prediction
using machine learning models under multiple feature selection
configurations. Three classifiers (CatBoost, Logistic Regression,
and Random Forest) were evaluated using four feature sets:
Full, No-LOS (excluding length of stay), Top-10, and Top-19.
SHAP-based feature importance was used to identify reduced and interpretable feature subsets. Performance was assessed using 5-fold cross-validation with precision, recall, F1-score, ROC-AUC, and PR-AUC. Experimental results show that CatBoost consistently achieved the best overall performance. The Top-19 feature set preserved nearly all predictive power of the Full model while reducing feature dimensionality by approximately 50%. In contrast, aggressive feature reduction (Top-10) and removing length of stay (LOS) resulted in statistically significant performance degradation. Paired t-tests further confirmed that CatBoost significantly outperformed Logistic Regression and Random Forest. These findings demonstrate that interpretable feature reduction combined with gradient boosting provides an efficient and effective approach for readmission prediction.
|
|||
| Keywords | hospital readmission prediction, machine learning, CatBoost, feature selection, SHAP, class imbalance, healthcare analytics, predictive modeling | |||
| Article #: RQD2026-21 | ||||
Proceedings of 31st ISSAT International Conference on Reliability & Quality in Design |