Expanded individual social determinants of health combined with clinical predictors in an XGBoost model improved 30-day heart failure readmission prediction (AUC 0.671 vs 0.632) and racial fairness.
Does the inclusion of expanded individual social determinants of health indicators improve the prediction of 30-day HF readmissions in patients with heart failure?
Including hundreds of individual social determinants of health indicators rather than traditional composite indices modestly improves machine learning models' ability to predict 30-day heart failure readmissions and enhances algorithmic fairness across racial groups.
Absolute Event Rate: 0% vs 0%
Background Heart failure (HF) hospitalization readmissions are associated with a high mortality rate and strain the health care system. Both clinical factors and social determinants of health (SDOH) predict HF readmissions, but the optimal approach to incorporating area‐level SDOH data remains unclear. Methods We merged census tract– and county‐level SDOH measures with electronic health record data in a retrospective cohort of 33 579 Black and White patients with HF (Emory Healthcare, 2010–2018). Six combinations of electronic health record data with 752 area‐level SDOH were evaluated using multiple machine learning models (logistic regression, random forest, XGBoost) to predict 30‐day HF readmission. Models were assessed for predictive performance using area under the receiver operating characteristic curve and algorithmic fairness across race using the equalized odds ratio. Results Expanded SDOH predictor sets improved predictive performance and algorithmic fairness compared with traditional SDOH indices. The XGBoost model using expanded SDOH alongside clinical predictors provided better predictive performance (area under the receiver operating characteristic curve, 0.671) and improved algorithmic fairness across patient race (equalized odds ratio, 0.437) than models with traditional indices (area under the receiver operating characteristic curve, 0.632; equalized odds ratio, 0.329). Feature‐importance analysis revealed that specific environmental predictors (housing cost burden, air quality) ranked among top predictors alongside clinical biomarkers. County‐level SDOH outperformed census tract–level SDOH and matched prediction performance of clinical predictors alone. Conclusions Including hundreds of individual SDOH indicators rather than traditional composite indices improves machine learning models' ability to predict 30‐day HF readmissions. While performance gains are modest, inclusion of specific environmental factors may provide greater clinical utility and improved equity across racial groups than composite SDOH measures for guiding further research and preventive interventions.
Fensore et al. (Sat,) reported a other. Expanded individual social determinants of health combined with clinical predictors in an XGBoost model improved 30-day heart failure readmission prediction (AUC 0.671 vs 0.632) and racial fairness.