Adaptation of a fair individualized polysocial risk score for hospitalization risk prediction

We adapted the individualized polysocial risk score (iPsRS), a machine learning model originally developed for patients with type 2 diabetes, to evaluate its generalizability in predicting 1-year hospitalization risk in a disease-agnostic adult cohort, with attention to fairness and explainability.

The study utilized de-identified electronic health record data from a retrospective cohort of 17 857 adult patients at the University of Florida Health. The original iPsRS framework was reimplemented using XGBoost and Logistic Regression models. These models were trained and fine-tuned using 13 individual-level social determinants of health (SDoH) predictors to predict all-cause hospitalization within 1 year. Models were fine-tuned at 5 levels (0%, 10%, 20%, 50%, and 70%) with 3 sampling strategies (none, oversampling, and undersampling). Model performance was primarily evaluated using the AUROC. Interpretability was assessed through SHapley Additive exPlanations and a causal structure learning algorithm, while fairness was analyzed by examining false negative rate disparities across age, sex, and racial/ethnic subgroups.

After fine-tuning, the adapted iPsRS demonstrated moderate predictive performance, achieving an AUROC of up to 0.671 with XGBoost and Normal sampling. SHapley Additive exPlanations analysis on the evaluation split showed age, food insecurity, marital status, and financial constraints emerged as consistently influential predictors. The causal structure analysis corroborated predictive findings, identifying age, race, and employment as proximal factors associated with hospitalization.

The iPsRS was successfully adapted beyond diabetes, maintaining moderate predictive performance and meaningful risk stratification. Incorporating individual-level SDoH enables equity-aware prediction, supporting broader use of iPsRS in clinical care.