158 - Integrating Social Determinants with EHR Risk Tools to Predict Medically Complex Pediatric Readmission

Publication Number: 3152.158

Arman Sandhu, Medical College of Wisconsin, Hoffman Estates, IL, United States; Ammar Amjad, Medical College of Wisconsin, Milwaukee, WI, United States; Omkar Chauhan, Medical College of Wisconsin, Milwaukee, NJ, United States; Yuvraj Singh, Medical College of Wisconsin, Milwaukee, WI, United States; Jessica Schnell, Medical College of Wisconsin, Milwaukee, WI, United States

Background: Unplanned pediatric hospital readmissions place emotional and financial strain on families and health systems, especially for medically complex children. Current prediction tools, such as the Epic readmission risk score, rely primarily on clinical data and often do not account for social determinants of health (SDoH) that affect a family's ability to manage care at home. Factors such as language barriers, transportation limits, caregiver stress, and insurance instability may elevate readmission risk even when clinical status appears stable. Incorporating SDoH into risk assessment may improve identification of patients needing enhanced discharge planning and follow-up support.

Objective: To compare the predictive performance of a custom SDoH-based risk score with the Epic readmission risk score in forecasting unplanned 30-day pediatric readmissions, and to determine whether a combined model enhances predictive accuracy.

Design/Methods: We performed a retrospective cohort study of 126 medically complex pediatric patients admitted to Children's Wisconsin between June and August 2025. For each patient, we recorded the Epic readmission score, calculated a custom SDoH score incorporating socioeconomic and caregiver-related factors, and identified 30-day unplanned readmissions. We examined score distributions, assessed Pearson correlation, generated receiver operating characteristic (ROC) curves, and calculated area under the curve (AUC). Sensitivity, specificity, and accuracy were evaluated using median score thresholds.

Results: The Epic and SDoH scores demonstrated moderate correlation (r = 0.52), indicating that each captures related but distinct risk dimensions. The Epic score achieved an AUC of 0.71, while the SDoH score achieved an AUC of 0.73, reflecting comparable predictive accuracy. The SDoH score showed higher sensitivity, identifying a greater proportion of true readmissions, though with more false positives. When combined in a logistic regression model, predictive performance improved further (AUC = 0.76), suggesting additive value in integrating clinical and social data.

Conclusion(s): SDoH factors contribute meaningfully to pediatric readmission risk and should be considered in prediction and care planning. The combined model demonstrated the strongest performance, highlighting the importance of incorporating social context alongside clinical information. Future work should validate this model in larger populations and evaluate targeted interventions triggered by elevated combined risk scores.

ROC Curves: Predicting 30 Day Readmissions
Receiver operating characteristic (ROC) curves for Epic (AUC = 0.71), SDoH (AUC = 0.73), and a combined logistic regression model (AUC = 0.76). The combined model demonstrated superior discrimination.

30 Day Readmission Counts
Overall 30-day readmission rates in the cohort (30% readmitted, 70% not readmitted)

Performance at Median Cutoffs (Epic ≥ 16, SDoH ≥ 3)
Sensitivity, specificity, and accuracy at median cutoffs (Epic ≥ 17, SDoH ≥ 3). SDoH thresholds yielded higher sensitivity but lower specificity compared to Epic.