Predicting Hospital Readmission in Medicaid Patients With Diabetes Using Administrative and Claims Data

Figure 1. Machine learning model development and validation workflow, showing stratified data splitting, hyperparameter optimization, cross-validation, final testing, SHAP explainability, and integration with the hybrid ML–PBPK digital twin framework.

Figure 2. Proposed digital twin–based clinical decision support architecture illustrating the integration of real-world patient data, machine learning risk prediction, PBPK-driven exposure simulation, and personalized clinical decision outputs.

Figure 3. Conceptual digital twin workflow showing parallel machine learning–based risk modeling and PBPK-based physiological simulation converging into an integrated personalized risk–exposure phenotype.

Figure 4. Age distribution of the inpatient encounters highlights a predominance of older adults, reflecting the typical demographic profile of hospitalized type 2 diabetes patients.

Figure 5. Class distribution showing an imbalance between readmitted and non-readmitted cases.

Figure 6. Distribution of admission types, emergency admissions dominate the cohort, indicating acute clinical presentations as a major driver of hospital encounters in diabetes patients.

Figure 7. Distribution of discharge disposition, patterns reveal variability in post-hospital care pathways, which are key determinants of readmission risk.

Figure 8. Distribution of insulin and metformin medication status. Medication adjustments during hospitalization are frequent and serve as indicators of disease instability and therapeutic response.

Figure 9. Training dynamics of the optimized XGBoost model. (a) Training and validation binary logistic loss across boosting iterations showing continuous reduction in loss during model optimization. (b) Training and validation AUC across boosting iterations. The validation AUC improves rapidly during the initial boosting iterations and reaches a plateau after approximately 50 iterations, whereas the training AUC continues to increase gradually until early stopping is triggered. This divergence indicates residual overfitting despite the application of regularization and validation-based early stopping, suggesting that additional boosting iterations primarily improved the model fit to the training data without corresponding gains in validation performance.

Figure 10. Stability of model performance across five-fold stratified cross-validation. Five-fold stratified cross-validation was performed exclusively on the 70% training subset during model development. Fold-wise AUC values demonstrate consistent predictive performance across different training partitions (mean AUC = 0.6793 ± 0.0072), indicating good model robustness and stability.

Figure 11. Predictive performance of the optimized XGBoost model on the independent test set. (a) Receiver Operating Characteristic (ROC) curve demonstrating an AUC of 0.686. (b) Precision–Recall (PR) curve illustrating classifier performance on the minority readmission class with an Average Precision (AP) of 0.221, providing complementary evaluation under class imbalance.

Figure 12. Calibration curve.