Assessing Equity Trade-offs in Risk Prediction Models When Using Social Determinants Instead of Race

• In Event: The Role of Big Data and Machine Learning in Health Care: Challenges and Opportunities

Abstract

Predictive models increasingly influence healthcare decisions by determining who receives care, interventions, and resources. As these tools become more central to clinical and policy workflows, a growing number of organizations have called for the removal of race as a predictor. A common proposal is to replace race with social determinants of health (SDoH), under the premise that SDoH better reflect structural inequities without invoking race itself. This shift is often motivated by concerns that including race risks reinforcing biological essentialism, the false belief that racial groups differ in innate biological ways. We strongly share this ethical concern. However, we argue that the current debate focuses too narrowly on conceptual harm and often overlooks a second important issue: the empirical consequences of removing race for model performance, fairness, and real-world impact.

These two concerns, ethical risk and practical outcomes, are distinct. Failing to evaluate the latter may lead to well-intentioned policies that unintentionally reduce equity and care quality. To support more evidence-based decision-making, we present an empirical case study using a machine learning pipeline that applies repeated cross-validation and resampling to compare the predictive performance and fairness of models. Both models were developed using data from the Coronary Artery Risk Development in Young Adults (CARDIA) study. One model includes race, while the other removes race and instead incorporates detailed SDoH, including financial strain, housing instability, access barriers, and experiences of discrimination. Because CARDIA provides exceptionally rich SDoH data that are rarely available in clinical workflows, this analysis offers an idealized scenario that illustrates both the promise and the limitations of replacing race with SDoH in practice.

We find a clear trade-off. Removing race and using SDoH reduces the number of individuals incorrectly flagged as high-risk. However, it also reduces the number of Black individuals who are correctly identified as being at high risk of cardiovascular disease. For every 1,000 Black individuals screened, the race-based model identifies six additional true cases but produces ninety-six more false positives. Missing true high-risk cases means that individuals who need preventive services or clinical interventions may not receive them. These differences have meaningful implications when applied across populations.

These findings point to a broader need for policy frameworks that explicitly assess the trade-offs involved in predictive model design. Whether models are used to guide treatment decisions, allocate preventive services, or determine reimbursement, the inclusion or exclusion of race is not a neutral choice. It directly affects who is flagged for care. Making these trade-offs explicit is a necessary step toward making informed decisions that align with values and priorities. These choices should be guided by context-specific evidence and made through inclusive, deliberative processes. Policymakers, regulators, payers, and health system leaders should engage communities, advocacy groups, and frontline professionals in evaluating these impacts. Making SDoH-based approaches viable will also require investment in data infrastructure to support the routine capture of high-quality SDoH. Our analysis illustrates how these assessments can be carried out and underscores the importance of balancing ethical commitments with practical consequences in advancing equitable healthcare.

Author

• Noah E Hammarlund, University of Florida
  Presenting Author