Explainability in Forecasting Hospital Readmissions with Machine Learning

Predictive modeling of healthcare needs to strike a balance between performance and interpretability-especially when used to guide decisions regarding resource allocation and patient risk. Here is a full machine learning workflow for predicting 30-day hospital readmission among diabetic patients based on the Diabetes 130-US Hospitals data. Following aggressive preprocessing, feature engineering, and deduplication, five classifiers were considered (logistic regression, decision trees, and ensemble learners such as Random Forest, XGBoost, and LightGBM). Although ensemble approaches had high accuracy (LightGBM: 91.4%), they had low recall because of readmission label imbalance (8.6%). To close the gap between performance and clinical confidence, we incorporated explainable AI (XAI) tools-SHAP for global and local model attributions, and LIME for case-by-case interpretability. SHAP explanations consistently pointed to length of stay, age, discharge disposition, and metabolic control (A1C values, glucose levels) as top drivers. LIME explanations served to distinguish true positives from missed cases and detected signal dilution in some medication or diagnostic patterns. In addition, we characterized high-risk patients based on SHAP-derived scores, demonstrating that they exhibited increased prior utilization, abnormal labs, and complicated discharge plans. Our framework illustrates how EHR-based classifiers enhanced with transparent XAI outputs can facilitate actionable, auditable, and clinically meaningful forecasting in the hospital environment. We publish reproducible pipelines and provide directions for next steps such as external validation and model threshold optimization for deployment-readiness.