Social Determinants of Health Improve Predictive Accuracy of Clinical Risk Models for Cardiovascular Hospitalization, Annual Cost, and Death

Background: Risk models in the private insurance setting may systematically underpredict in the socially disadvantaged. In this study, we sought to determine if US minority Medicare beneficiaries had disproportionately low costs compared to their clinical outcomes and if adding social determinants of health (SDOH) into risk prediction models improves prediction accuracy.

Methods and Results: Retrospective observational cohort study of 2016-2017 Medicare Current Beneficiary Survey (MCBS) data (N=3,614) linked to Medicare fee-for-service claims. Logistic and linear regressions were used to determine the relationship between race/ethnicity and annual costs of care, all-cause hospitalization, CV hospitalization, and death. We calculated observed to expected (O:E) ratios for all outcomes under four risk models: 1) age + gender 2) Model 1 + clinical comorbidity adjustment, 3) Model 2 + SDOH, and 4) SDOH alone. Our sample was 44% male and 11% Black or Hispanic. Among minorities, adverse clinical outcomes were inversely related to cost. After multivariable adjustment, Blacks/Hispanics had higher rates of CV hospitalization (IRR 1.78, p=0.012), but similar annual costs ($-336, p=0.77) compared to whites. Among whites, models 1-4 all showed similar O:E ratios, suggesting high accuracy in risk prediction using current models. Among minorities, adjustment for age, gender, and comorbidities under-predicted all-cause hospitalization by 20% (O:E=1.20), CV hospitalization by 70% (O:E=1.70); and overpredicted death by 21% (O:E=0.79); adding SDOH brought O:E near 1 for all outcomes. Among both groups, the SDOH risk model alone performed with equal or superior accuracy to the model based on clinical comorbidities.

Conclusions: A paradoxical relationship was observed between clinical outcomes and costs among racial/ethnic minorities. Because of systematic differences in access to care, cost may not be an appropriate surrogate for predicting clinical risk among vulnerable populations. Adjustment for SDOH improves the accuracy of risk models among racial/ethnic minorities, and could guide use of prevention strategies.