Algorithms play an important role in our daily lives — they unlock our phones, tell us which movies to watch, and decide what we see on social media.Healthcare providers also often use these calculations to inform diagnosis and treatment plans, so it’s important to remember that algorithms are only as good as data for training them.
Much of the data used to train healthcare algorithms today reflects the structural racism that exists in the U.S. healthcare system, creating biases that negatively impact the health outcomes of already marginalized populations. That’s why engineers building healthcare algorithms should move away from using race as a measure of health disparities and instead use social determinants of health. control panel On medical AI bias.
“There is bias in the world. If it is in the world, it is in the data,” Seun Ross said. Independence Blue Cross‘ Executive Director of Health Equity.
When such biased data is used to train healthcare algorithms, structural racism in the healthcare industry perpetuates. To illustrate this point, Ross cites an example of racist bias in the history of the American healthcare system: When black and white patients exhibited the same behavioral health symptoms, black patients were more likely to be diagnosed with schizophrenia , while white patients were more likely to be diagnosed with schizophrenia. More likely to be diagnosed with mood disorders such as depression and anxiety. If an engineer trained a machine-learning program using the historical health records of black patients treated for schizophrenia in the U.S., it could arguably use racist predictors to make calculations and continue to overdiagnose black patients with schizophrenia, Ross said.
Another panelist agreed.
Algorithms will continue to retain the racial biases that have historically been baked into the U.S. health care system unless the medical field stops enforcing racial discrimination in its machine learning models, said Dr. Jaya Aysola, assistant dean for inclusion and diversity at the Perelman School of the University of Pennsylvania. medicine.
Dr. Aysola co-authored a article last year in New England Journal of Medicine AnnounceRing races are not biological categories that produce unequal health outcomes based on innate differences, but social categories that reflect the impact of unequal social experiences on health. She and her co-authors argue that medical education and practice fail to reflect these advances in understanding the relationship between race, racism and health.
“If we start to see health disparities between races, but we know it’s not inherent in genetic susceptibility, lack of individual agency, or anything inherent in that population of individuals, we have to ask the next question: So, what’s the difference? What does race represent?,” Dr. Isola said. “When we started looking at this, we realized that race was really the proxy for structural racism. ”
According to Drs Ross and Isola, the idea that someone’s race predisposes them to a certain health condition is a fundamental misunderstanding because race is socially constructed.
Genetic ancestry — which is a biological trait that is distinct from ethnicity — can sometimes affect a patient’s likelihood of being diagnosed with a disease. But the real variable that a machine learning model should take into account is the unequal treatment patients receive due to structural racism, which often has a major impact on their health outcomes.
“I think we should use social determinants of health instead of race as a measure of health disparities in research guidelines and standards of care,” Ross said. “Factors such as whether you have health insurance, your community, education, occupation, income, wealth, health behaviors, sexual identity, gender, disability status, age and how you experience racism are more important factors.”
Photo: syahrir maulana, Getty Images
