Racial Bias in Medical Care Decision-Making Tools

These tools contain algorithms, or step-by-step procedures, usually computerized, for calculating factors such as risk of heart disease, the need for a chest X-ray, and prescription medicine dosage. Artificial intelligence can be used to scour health records and billing systems to create the needed data sets.

On the surface, these factors all sound very objective. But recent studies have shown that the data analysis used in these algorithms can be biased in crucial ways against certain racial and socioeconomic groups. This can have myriad consequences in terms of the amount and quality of healthcare that people in these groups receive.

Racial Bias Affects the Sickest Patients

In 2019, a study of an algorithm widely used by U.S. hospitals and insurers to allocate extra health management assistance was shown to systematically discriminate against Black people.1 The decision tool was less likely to refer Black people than White people to care-management programs for complex medical needs when both racial groups were equally sick.

The underlying reason for the bias was linked to the algorithm’s assignment of risk scores to patients based on their previous year’s medical costs. The assumption was that identifying patients with higher costs would identify those with the highest medical needs.² However, many Black patients have less access to, less ability to pay for, and less trust in medical care than White people who are equally sick. In this instance, their lower medical costs did not accurately predict their health status.3

Care-management programs use a high-touch approach, such as phone calls, home visits by nurses, and prioritizing doctor appointments to address the complex needs of the sickest patients. The programs have been shown to improve outcomes, decrease emergency room visits and hospitalizations, and decrease medical costs. Because the programs themselves are expensive, they are assigned to people with the highest risk scores. Scoring techniques that discriminate against the sickest Black patients for this care may be a significant factor in their increased risk of death from many diseases.3

Race as a Variable in Kidney Disease

Algorithms can contain bias without including race as a variable, but some tools deliberately use race as a criterion. Take the eGFR score, which rates kidney health and is used to determine who needs a kidney transplant. In a 1999 study that set the eGFR score criteria, researchers noticed that Black people had, on average, higher levels of creatinine (a byproduct of muscle breakdown) than White people did. The scientists assumed that the higher levels were due to higher muscle mass in Blacks. They therefore adjusted the scoring, which essentially meant that Black people must have a lower eGFR score than Whites to be diagnosed with end-stage kidney disease. As a consequence, Blacks have had to wait until their kidney disease reached a more severe stage in order to qualify for treatment.

Recently, a student of medicine and public health at the University of Washington School of Medicine in Seattle observed that eGFR scores were not accurate for diagnosing the severity of kidney disease in Black patients. She fought to have race removed from the algorithm, and won. UW Medicine agreed that the use of race was an ineffective variable and did not meet scientific rigor in medical diagnostic tools.4

Body Mass Index and Bias

Even the simplest medical decision tool that does not include race can reflect social bias. The body mass index (BMI), for example, is based on a calculation that multiplies weight by height. It is used to identify underweight, overweight, and obese patients.

In 1985, the National Institutes of Health tied the definition of obesity to an individual’s BMI, and in 1998 an expert panel put in place guidelines based on BMI that moved 29 million Americans who had previously been classified as normal weight or just overweight into the overweight and obese categories.6 By BMI standards, the majority of Blacks, Hispanics, and White people are now overweight or obese. The 2018 percentages for obesity are roughly equal for Black, Hispanic, and white men (ranging from 31.2% to 34.2%). But the percentages of women who are labeled obese by BMI are:7 8

• 44.2%—Black
• 35.4%—Hispanic
• 28.7%—White

An atmosphere of weight-shaming and distrust

Branding such large percentages of populations as overweight or obese has created an atmosphere of weight-shaming and distrust between patients and doctors. Higher-weight people complain that doctors don’t address the health problems or concerns that brought them in for a checkup. Instead, doctors blame the patient’s weight for their health issues and push weight loss as the solution. This contributes to Black and Hispanic patients avoiding healthcare practitioners and thus perhaps missing opportunities to prevent problems or catch them early.

Furthermore, it is becoming increasingly clear that being overweight or obese is not always a health problem. Rates for some serious conditions, such as hospitalization for COVID-19, high blood pressure, heart disease, stroke, type 2 diabetes, and other diseases, are higher among those who are obese.⁹  But for other conditions—such as recovery from serious injury, cancer, and heart surgery—higher weight people have better survival rates.10 11 12

New, improved Canadian guidelines

In fact, new obesity guidelines for Canadian clinicians, published in August 2020, emphasize that doctors should stop relying on BMI alone in diagnosing patients. People should be diagnosed as obese only if their body weight affects their physical health or mental wellbeing, according to the new guidelines. Treatment should be holistic and not solely target weight loss. The guidelines also note that: “People living with obesity face substantial bias and stigma, which contribute to increased morbidity and mortality independent of weight or body mass index.»13

Reducing Bias in Decision Tools

Medical algorithms are not the only type of algorithm that can be biased. In 2018, for example, Amazon stopped using a recruitment tool that showed bias against women. The tool, which analyzed 10 years of hiring data during a period when Amazon had predominantly hired men, had used that history to teach itself to prefer male candidates.14

In healthcare, machine learning often relies on electronic health records. Poor and minority patients may receive fractured care and be seen at multiple institutions. They are more likely to be seen in teaching clinics where data input or clinical reasoning may be less accurate. And patients may not be able to access online patient portals and document outcomes. As a result, the records of these patients may have missing or erroneous data. The algorithms that drive machine learning may thus end up excluding poor and minority patients from the data sets and needed care.15

The good news is that awareness of biases in healthcare algorithms has grown in the past few years. Data input and outcomes are being checked for racial, ethnic, income, gender, and age bias. When disparities are recognized, the algorithms and data sets can be revised toward better objectivity.