The use of race in clinical prediction models (CPMs) is one of the most contentious issues in modern healthcare. While "race-blind" approaches aim for neutrality, they often inadvertently mask systemic inequities, leading to inaccurate risk assessments for minoritized populations. Conversely, "race-aware" models are criticized for potentially reifying biological myths of race. To navigate this complexity, the GUIDE (Guidance for Unbiased predictive Information for healthcare Decision-making and Equity) framework introduces 31 consensus-based recommendations for the ethical use of race as a prognostic factor.
Foundational Principles
The GUIDE framework is built upon several core premises that redefine how developers should view demographic variables. Central to these is the recognition that race is a social construct rather than a biological one. In the context of predictive modeling, race serves as a proxy for unmeasured social determinants of health, environmental exposures, and the cumulative physiological toll of systemic racism.
A critical distinction made by the framework is between prediction and causal inference. While a variable like race may be a powerful predictor of an outcome due to societal correlations, its inclusion does not imply a causal mechanism. Understanding this allows developers to use race to improve model accuracy without making unfounded biological claims.
The Polarity of Context
GUIDE moves beyond a one-size-fits-all approach by categorizing clinical decisions into "polar" and "non-polar" contexts:
- Non-Polar Contexts: These involve shared decision-making (e.g., whether to undergo a screening test). Here, the priority is maximum predictive accuracy for the individual patient.
- Polar Contexts: These involve the allocation of scarce resources (e.g., organ transplant lists). In these scenarios, distributive justice and fairness are the primary concerns, and race should only be included if it actively reduces existing disparities.
Methodology
The framework was developed through a rigorous 5-round modified Delphi process involving a 14-person Technical Expert Panel (TEP). This multidisciplinary group—comprising clinicians, computer scientists, ethicists, and statisticians—required a 75% supermajority for each item. Crucially, the process incorporated patient focus groups to ensure the guidelines reflected the values of those most impacted by algorithmic decisions.
Impact and Results: The Case for Calibration
A key contribution of GUIDE is the emphasis on calibration—ensuring that predicted risks match observed outcomes across all racial subgroups—rather than relying solely on aggregate metrics like AUC.
To illustrate this, the authors applied the framework to a dataset of 1.1 million prediabetic patients from the OptumLabs Data Warehouse. A "race-blind" model was found to systematically underestimate diabetes risk in Black patients. By transitioning to a "race-aware" model, the proportion of Black patients correctly identified as high-risk increased from 31% to 51%, significantly improving the model's clinical utility for this subgroup.
Conclusion
The GUIDE framework provides a necessary roadmap for the transparent and ethical use of race in clinical algorithms. While currently focused on the U.S. landscape, it sets a global precedent for moving toward "antisubordination" in healthcare technology. As a living document, GUIDE will continue to evolve alongside advancements in machine learning and our societal understanding of equity, ultimately aiming for a future where high-quality data on social determinants of health can supplement or replace race as a proxy.
Summary updated