Last year, three inflamed, welt-like bumps appeared on my two-year-old son’s shoulder, arm, and temple. So I called the pediatrician’s office in my rural town.
The doctor seemed unconcerned and prescribed an anti-bacterial and a steroid cream. Days later, the bumps multiplied, covering my son’s torso, arms, and legs. A different doctor saw him, asked about our family’s history of skin conditions, and gave a diagnosis that left me confused: atopic dermatitis, commonly known as eczema.
Her fast, perfunctory conclusion seemed way off. Eczema is a chronic condition; my son’s rash looked different and came on suddenly. Doubt filled my mind: Was the doctor simply aware that eczema is 1.7 times more common in black children than white ones? Was my experience an example of a persistent problem in medical diagnosis — data-driven assumptions, based on flawed ideas about race?
Like many other industries, medicine is increasingly turning to artificial intelligence in an effort to improve outcomes. The FDA recently approved a machine-learning algorithm to measure blood flow through the heart. Doctors can now use computers to diagnose patients based on images, physician notes, and electronic health records.
We trust these algorithms because we assume that the data they use is evidence-based and representative of society, says LaVera Crawley, who was an empirical bioethicist at Stanford University for more than 16 years. But “algorithms are programs, which means they are programmed by humans, which means human bias can show up in the algorithm,” she says.
The source of the data is one major risk factor, she says — particularly when it comes to race. A lot of aggregated medical data stems from clinical trials, which often have low levels of participation from racial minorities and other marginalized groups.
Data blind spots can have dangerous implications, Crawley points out.
For example, self-driving cars have been found to be more likely to hit people with darker complexions. Data-driven tools used to predict future criminal behavior can unfairly impact people of color. Likewise, incomplete medical data could increase health inequality. Conditions like skin cancer, which present differently and with higher mortality in darker-pigmented patients, are harder for machines to detect if their calculations are based on “typical diagnostic factors.”
“It’s like trying to go get your DNA tested to find out where you are from in Africa and going to a company that doesn’t have that gene pool represented,” Crawley says. “You’re not going to get accurate information.”
The roots of the data deficit are complex and fraught. But one of them is rational distrust. Crawley points to the U.S. Public Health Service’s “Tuskegee Experiment” — in which 399 black men with syphilis were studied from 1932 to 1972 but never offered proper treatment — and the unauthorized 1951 harvesting of patient Henrietta Lacks’ cancer cells, which are still used in cancer research.
“African Americans are considered mistrustful of the health care system, as if it’s our disposition,” says Crawley. “But instead, it’s the situations that we find ourselves in that make us not trust.”
And while we discuss racial bias in medicine as if it’s a thing of the past, it’s also a part of the present. A 2019 study in Science Daily found that an overwhelming number of black Americans with severe depression are misdiagnosed as schizophrenic. A 2016 study from the National Academy of Sciences found that half of white medical students and residents surveyed “hold false beliefs about biological differences between blacks and whites,” which cause them to underestimate black patients’ pain, downplaying or downright ignoring their complaints.
“Many well-researched studies show that black people are under-treated for pain,” says Carolyn Roberts, a Yale University professor of history of science and medicine and African American studies. “That implicit bias results in misdiagnoses and substandard treatment.”
“It’s like trying to go get your DNA tested to find out where you are from in Africa and going to a company that doesn’t have that gene pool represented. You’re not going to get accurate information.”
ProPublica’s 2017 project “Lost Mothers” personalized the epidemic of maternal deaths in the United States — a crisis even more pronounced among black women. I could have been one of the women in that story: After my son was born, I spent a month insisting on treatment for what turned out to be a potentially life-threatening condition.
Armed with this knowledge and personal history, I was concerned about the likelihood my son would be misdiagnosed — and knew that growing up black increases his odds. Some health professionals worry that the problem of misdiagnosis will get worse if doctors increasingly rely on population data to make their diagnoses.
“Diagnostic shortcuts minimize the fact that race is a social construct, not a biological one,” says Jaime Slaughter-Acey, a professor of epidemiology and community health at the University of Minnesota. “Implicit biases, stereotypes, and prejudices can be inserted consciously or unconsciously.”
And sometimes it’s not the data itself, but the way the data is used, that can put patients at risk, Crawley says.
“The way you use population data is to allow yourself to develop a hypothesis,” she says. “But the requirement is you test that hypothesis instead of just acting on it. When you act on it without testing it out, that’s being biased.”
Lately, some medical systems and advocacy groups have focused intently on this challenge, trying to help doctors and nurses use health data to improve outcomes, rather than further bias.
One solution is to check existing data with more data. Ava Liberman, a professor of neurology at Albert Einstein College of Medicine, recommends an approach in which analysts use administrative claims data to measure diagnostic errors and learn how they connect to race, sex, and other patient characteristics.
“Measuring diagnostic errors will improve diagnostic safety for everyone, including marginalized groups,” says Liberman.
But to truly reduce bias in medicine — and build trust among patients — others say the medical profession needs to change its very makeup. That means recruiting more participation from minority groups in medical studies and educating and hiring more medical professionals of color. (One randomized clinical trial, published in the National Bureau of Economic Research in October 2018, found that black men receive more effective care from black doctors.)
Today, only about 6 percent of medical school graduates are black, while only 7 percent are Hispanic. Medical school faculty, who transmit their personal experience to future professionals, are similarly lacking in diversity. Organizations like Tour for Diversity in Medicine, known as T4D, are working to change the balance, sending physicians, dentists, pharmacists, and medical students of color to college campuses to demonstrate to students that they can have a future in medicine.
Meanwhile, Slaughter-Acey says all clinicians need to work harder to earn their minority patients’ trust. She recommends that they practice “cultural humility,” which she defines as “a lifelong commitment that involves self-evaluation and self-critique when learning about another person’s culture.”
More trusting relationships, Slaughter-Acey says, would improve communication and care — encouraging doctors to look beyond statistics when they’re interacting with patients, and encouraging patients to talk more openly about their conditions.
When it came to my son’s case, I know trust was a major barrier. His rash eventually faded, with the mystery never solved. And there’s certainly a chance that my motherly instincts were wrong about his diagnosis. My journey to parenthood didn’t come with a M.D.
But through the years, I’ve had many first-hand encounters with race-based assumptions in medicine. One nearly killed me. So I’ve never stopped approaching medicine with skepticism — or thinking that data, useful as it seems, can be as biased as a person in the room.