Leveraging a machine learning model to predict hospital readmission risk: integrating clinical and social determinants of health data.

Abstract

Background: Hospital readmissions remain a major challenge for healthcare systems, contributing to higher costs and worse patient outcomes. Although most prediction models rely primarily on clinical data, integrating social determinants of health (SDOH) may improve risk assessment. However, the use of machine learning (ML) to combine clinical and SDOH data for readmission prediction remains limited.

Objective: To develop and compare machine learning models for predicting 30-day hospital readmission by integrating clinical and SDOH data.

Methods: We conducted a retrospective cohort study of 3,018 adult patients discharged from a large academic medical center between January 2022 and December 2023. Clinical variables were extracted from electronic health records and linked, through geocoded residential addresses, to area-level SDOH indicators from publicly available census data, including neighborhood deprivation, median income, and educational attainment. Six tabular ML models were trained and evaluated, including Logistic Regression, Random Forest, XGBoost, LightGBM, CatBoost, and Support Vector Machine. Model performance was assessed using the area under the receiver operating characteristic curve (ROC-AUC), precision-recall AUC (PR-AUC), and F1-score. SHapley Additive exPlanations (SHAP) were used to assess feature importance.

Results: Ensemble models outperformed Logistic Regression, with XGBoost achieving the best performance on the test set (ROC-AUC 0.79, 95% CI 0.75-0.82; PR-AUC 0.71). In addition to key clinical variables such as prior admissions and comorbidity burden, SDOH features including neighborhood socioeconomic status and household composition were among the most important predictors.

Conclusion: Integrating clinical and SDOH data into ML models improved prediction of 30-day hospital readmission. These findings support moving beyond clinical-only models and suggest that SDOH-informed prediction may help identify high-risk patients earlier and guide more targeted care management.