Researchers wanted to see how different ways of calculating heart disease risk affect patients. They studied over 3,200 Black and White adults who did not have known heart disease at the start. The team compared three prediction models: one that included race, one that replaced race with social factors like education and income, and one that used only clinical information like blood pressure and cholesterol. They tracked who developed heart disease over about a decade.
All three models performed similarly in their overall ability to predict who would get heart disease. However, the models that did not use race had important trade-offs. The model using social factors improved some fairness scores but systematically underestimated risk, leading to more Black participants being recommended for treatment when they might not have needed it. The model using only clinical data also improved fairness but missed treating four people who later developed heart disease, all of whom were Black.
The study shows that simply removing race from a risk calculator does not automatically fix fairness problems and can create new ones. The main reason to be careful is that no single measure of fairness or accuracy captured the full picture of how these models would affect different groups of people. Readers should understand that this was a look back at existing data, not a test of new tools in real-time. Health systems will need thorough testing before they can be confident that changing their risk models truly helps those most at risk.