Machine learning-driven risk prediction for post-hospitalization diabetes case management: Integrating clinical and social determinants of health.

Abstract

Patients hospitalized for diabetes-related conditions face elevated risks of emergency department (ED) visits post-discharge, driven by both clinical factors and social determinants of health (SDoH). This study aimed to develop and validate predictive models integrating clinical and SDoH data to identify high-risk patients for post-hospitalization diabetes case management. We conducted a retrospective cohort study using electronic health record data from the University of Alabama at Birmingham Medical Center, including 162,063 inpatient encounters (January 2020-June 2024) for training and testing and 16,164 encounters (January-May 2025) for temporal validation. Patients were identified by diabetes-related ICD-10 codes or HbA1c ≥ 6.5 %, reflecting the scope of the institution's diabetes case management program. Predictors included demographics, diabetes-related comorbidities, surgical procedures, laboratory values, medications, and both area-level and individual-level SDoH. Logistic regression, decision trees, and XGBoost models were developed to predict diabetes-related ED visits within 3 months post-hospitalization. Hyperparameters for decision tree and XGBoost models were tuned via 10-fold cross-validation, and calibration was assessed using Brier scores and calibration plots. Among 162,063 hospitalizations, 6.2 % resulted in a diabetes-related ED visit. XGBoost achieved the best performance (area under the curve [AUC] 0.846, precision 0.420, sensitivity 0.296, specificity 0.972), maintained on temporal validation (AUC 0.842). Key predictors included past ED visit frequency, insulin prescriptions, age, and area-level SDoH indices. Individual-level SDoH factors, including home safety issues and work disability, also contributed to prediction. Targeting the top 20 % of predicted risk captured 64.1 % of all ED visits. Model discrimination was consistent across racial subgroups (AUC range: 0.843-0.851). Calibration was clinically acceptable across datasets. Integration of clinical and SDoH data achieved effective prediction of post-hospitalization ED visits. XGBoost provided excellent discrimination with temporal stability. Decision trees offered greater interpretability. A pilot implementation delivering daily risk-stratified patient lists to the diabetes case manager is underway, demonstrating a practical pathway from model development to clinical decision support.