Can Machine Learning Eliminate America’s Racial Disparities in Medicine?
Ask almost any Black American seeking medical assistance and they have at least one story about a time they’ve received subpar care from a physician. In fact, unless they have a dedicated, and usually Black, general physician, chances are most experiences they’ve had are negative. While the issue has been around for hundreds of years, thanks to platforms like Tiktok, more physicians are openly acknowledging medicine‘s long history of racism.
Medical racism heavily contributes to the high rates of maternal mortality, extremely disproportionate rates of Covid-19 infection and death, and the lowest rates of receiving proper healthcare from the same doctors and hospitals as the other demographics of the general population. In fact, in a 2016 survey assessing racist beliefs of medical students and residents, at least half of those surveyed believed racist medical myths, such as Black people: having thicker skin; feeling less pain because we have fewer/less sensitive nerve endings; and having extra biological and anatomical structures than White people. While medical racism should be at the forefront of the discussion as long as it continues to exist, it is especially pertinent during the COVID-19 pandemic.
Disproportionate care & high mortality rates
In Chicago, alone, 70% of the people who suffered COVID-19 related deaths were Black, despite only 30% of Chicago’s overall population being Black. COVID-19 has been particularly difficult in Indian Country, spiking hospitalization to 5.3 times higher than White counterparts. While many factors, such as socioeconomic standing, have been suggested contributions to these numbers, the largest contribution was found nearly 20 years ago: Black and Brown people simply receive a lower quality of care.
According to “Unequal Treatment: Confronting Racial and Ethnic Disparities in Health Care,” a 2003 report conducted by the Institute of Medicine, now the National Academy of Medicine, socioeconomic class, insurance, access to healthcare, age, and other factors could not account for the differences in healthcare. When comparing that level of care to White counterparts with similar income, age, conditions, doctors, and severity, researchers found a significant pattern of “concrete, inferior care that physicians give their black patients.”
From global pandemics to the highest contributing factors of death, why racial and ethnic minorities receive a lower quality of care is complicated, but inadequate academic and professional license is strongest amongst them. To further ingrain racial disparities in medicine, examples of most illnesses with outward symptoms, such as signs of melanoma, are almost exclusively shown on white-skinned or slightly tanned bodies, contributing to a 65% rate of survival rates for Black patients compared to a 91% rate for White patients. For greater context, 1 in 38 White people have melanoma compared to 1 in 1000 Black people.
Is machine learning the answer?
The question is: can machine learning successfully close the racial disparities in medicine for Black patients and other patient of color? It depends, but so far the answer is a resounding “NO!”
Unfortunately, machine learning models are subject to the same biases the modeler and the data hold. In fact, millions of Black patients are already victim to bias health-care algorithms in a system that already leaves Black patients with larger medical debt in exchange for lower quality care. While applying machine learning models seems like a great way to improve the quality in care amongst all races, the same racist myths and tendencies that infect medicine also infect data science. Computer vision carries over legacy artifacts from its photography forefather which rarely takes in consideration non-White skin tones. Recommendation models are only as good as the quality and accuracy of data used to train and test it. Image classification models have relied heavily on non-Black patients, so existing models fail to consider overlapping health conditions that disproportionally affect Black patients and contribute to higher rates of misdiagnosis.
One required solution is to increase racial diversity; not only in data but in modelers, too. Only about 5% of software engineers are Black, though the non-Latinx Black population accounts for more than 13% of the national population, with an even lower percentage for Black data modelers and scientists. Comparatively, Latinx engineers, of any race, make up only 6% of the industry, even though most tech hubs are in cities with high Latinx populations. It’s a common practice that Black users or audiences are rarely considered in the development of AI and healthcare products, even in issues that mostly affect Black people. If more Black modelers were included from inception, in comparison to models made by largely White and Asian teams, we could eliminate much more bias before models even launch. And while more diverse data teams might breakdown biases built into the foundation of these models, machine learning still isn’t end all of medical racism.
These racist myths didn’t simply show up in 2016. They’ve existed for nearly 400 years in the U.S. and are still being taught by medical schools, residents, attendings, and practitioners. Even with a perfect model infrastructure, the model’s only as good as the data used for testing, training, and validation. While technology can help us improve healthcare, if medicine doesn’t change for the better, we’ll only contribute to centuries-long disparities and risk putting countless millions in danger.
