Racial Bias in Healthcare Algorithms

Racial Bias in Healthcare Algorithms

A recent study in Science found that a commercial algorithm used to guide healthcare decisions operated with heavy racial bias. The algorithm assigns risk scores to patients so physicians can determine which patients may need more specialized medical care. When comparing scores assigned to Black and White patients who were equally ill, researchers found that Black patients' scores were consistently lower than their White counterparts'. This meant that they were less likely to be reffered to more-personalized care programs, compromising their quality of care compared to White patients.

This bias occured because the algorithm determined patient risk based on their health care expenditure in one year. While it is true that higher healthcare spending is correlated with higher healthcare needs, systemic racism's effects, including Black patients' distrust of the healthcare system and physician implicit biases result in Black having lower access to care, reducing their spending. On average, White patients recieved $1,800 more care compared to Black patients with the same number of chronic health conditions.

Only 17.7% of patients that the algorithm assigned to receive extra care were black. The researchers calculate that the proportion would be 46.5% if the algorithm were unbiased.

Dr. Ziad Obermeyer, an Associate Professor (Acting) at the University of California, Berkeley, identified factors, other than health care spending, that could indicate patient health risk and found than an algorithm with these changes reduced bias by 84%.

As healthcare becomes more and more automated, it is important to identify and fix issues of bias and disparities. While the algorithm itself cannot be inherently racist, building it to utilize parameters that are affected by structural racism and bias, can be incredibly damaging. At Caralyst, we are fighting against structural racism in data science by not using proxies or making assumptions about patients' needs based on their demographics. Rather, our matching algorithm is solely based on patients' needs.