Beyond Composite Indices: Comprehensive Social Determinants Improve Heart Failure Readmission Prediction.

Area of Science:

• Health services research
• Machine learning in healthcare
• Health equity

Background:

• Heart failure (HF) readmissions carry high mortality and healthcare costs.
• Predicting HF readmissions involves clinical factors and social determinants of health (SDOH).
• Optimal use of area-level SDOH data for prediction is not well-defined.

Purpose of the Study:

• To evaluate the impact of incorporating extensive area-level social determinants of health (SDOH) measures into machine learning models for predicting 30-day heart failure (HF) readmissions.
• To assess the predictive performance and algorithmic fairness across racial groups of different SDOH data combinations.

Main Methods:

• Retrospective cohort study of 33,579 Black and White patients with HF.
• Merged electronic health record data with 752 area-level SDOH measures (census tract and county).
• Evaluated six combinations of SDOH and EHR data using logistic regression, random forest, and XGBoost models; assessed using AUC and equalized odds ratio.

Main Results:

• Expanded SDOH predictor sets enhanced predictive performance and algorithmic fairness compared to traditional SDOH indices.
• The XGBoost model with expanded SDOH achieved an AUC of 0.671 and an equalized odds ratio of 0.437.
• Specific environmental SDOH (e.g., housing cost burden, air quality) were key predictors; county-level SDOH matched clinical predictor performance.

Conclusions:

• Utilizing hundreds of individual SDOH indicators, rather than composite indices, improves machine learning model prediction of 30-day HF readmissions.
• Inclusion of specific environmental factors offers potential for greater clinical utility and improved racial equity in interventions.
• Findings support the use of detailed SDOH data for refining predictive models and guiding preventive strategies for HF readmissions.