How Bias Creeps Into Health Data. What Can Be Done?

As data transforms healthcare, hidden biases can undermine equity

While digital health solutions offer great promise for improving clinical outcomes and care access, racial biases embedded in some of these technologies and data algorithms could undermine health equity efforts. Inequities in overall technology access can also exacerbate existing health equities for historically marginalized people. As health data proliferates, identifying and understanding these often hidden biases are key in advancing health equity efforts.

What is digital health?

Digital health refers to any interactions with the health system that involves technology. These can include everything from mobile well-being apps to wearable devices and telehealth appointments. The term also encompasses data-driven healthcare predictions and artificial intelligence guidance that is increasingly used in clinical decision making.

As technology becomes more widespread, equity advocates are increasingly raising concerns that bias—even unintentional— may be embedded in these human-created systems. These can range from racially-biased data to inequities in access to the Internet and can widen already persistent health inequities for historically marginalized people.

“Digital back doors" can jeopardize health

Much like the literal back doors that Black people and other people of color were once forced to use to enter restaurants and other public spaces, these "digital back doors" “may once again make them and others second-class citizens when it comes to health," Brown professor Kim Gallon writes in this STAT article.

“By jeopardizing the health of patients of color, digital back doors are a feature of health inequity," Gallon added in a recent interview with the National Library of Medicine's "Circulating Now."

How human biases can impact health data

Bias can arise even in seemingly impartial data predictions. In recent years, there have been multiple examples of flawed algorithms used for medical decision making that negatively impacted the quality of care for historically disadvantaged people. As this STAT article points out, the origins can often be traced back to outdated science or human-biased data, such as assuming race was the cause of poor health outcomes.

Multiple studies demonstrate flawed algorithms

In 2020, a widely-cited study published in the New England Journal of Medicine found that implicit racial bias was tainting algorithms driving care decisions across multiple medical specialties, from heart surgery to obstetrics.

In another well-reported example, an algorithm that took race into account in establishing kidney function potentially made Black people's kidneys seem healthier than they were, according to the findings published in JAMA Network. Removing race would have raised the number of Black adults eligible for a kidney transplant wait list.

Addressing race in statistical analysis is fraught with potential issues. Although often done, controlling or adjusting for race can lead to misestimating the total impact of interventions while simultaneously erasing the effect in the non-reference group. When setting defaults for the reference group to White people, one can easily “control away" issues more frequently experienced in non-White populations by naively controlling for race variables.

Cases where we are interested in differences in impact between groups are often performed inappropriately. Just the choice between interaction vs. stratification by race can lead to substantial methodological and interpretation issues, as discussed in more detail here. These methodological issues are all in addition to and compounded by social issues related to social biases.

Using proxies has health equity consequences

Yet another example of biased algorithms came from a formula used to predict patients that required additional healthcare services such as home visits, according to research published in Science. In this case, an algorithm used one's medical history to predict their future health care costs.

“But cost is not a race-blind metric: for socioeconomic and other reasons, black patients have historically incurred lower health-care costs than white patients with the same conditions. As a result, the algorithm gave white patients the same scores as black patients who were significantly sicker," explains Charlotte Jee in the MIT Technology Review article on the technology flaw.

Stigmatizing language in electronic records

The impact of healthcare digitization on health equity isn't limited to data and technology access. There can be direct personal biases that work their way into electronic records, documents that can follow a patient through future encounters with the healthcare system. Clinicians were more likely to use “stigmatizing" language in electronic notes about Black patients compared to non-Hispanic White patients, according to a 2022 study published in JAMA Network Open.

“These findings suggest that improved conscientiousness and training around avoiding stigmatizing language in medical notes could improve health equity," wrote the study's authors.

Unequal access to high-speed internet

Even as the Internet becomes seemingly ubiquitous, people living in the United States have unequal access to high-speed broadband. As technology becomes more integrated in healthcare setting, these gaps can become especially problematic for accessing care.

About a quarter of US adults with household incomes below $30,000 a year report not owning a smartphone, according to 2021 Pew Research Center findings. About 40% of adults with lower incomes don't have broadband services at home or a desktop/laptop computer. Those findings are in sharp contrast to higher earners who are more likely to have multiple devices that can access the Internet.

Along with these socioeconomic disparities in access, discrimination from Internet service providers can add to the access challenges. Much like the discriminatory practices that restricted access to home loans and ownership for Black people, the practice of limiting Internet access or upgrades in certain regions can impact who can access digital healthcare.

These inequities became painfully evident during the pandemic when people living in areas with good high-speed Internet were more easily able to communicate with physicians via telehealth visits.

Data has an important role in advancing health equity

While flawed data and biased structures can enhance inequities, that doesn't necessarily mean we shouldn't use race-related data. In fact, in some cases, collecting race and ethnicity-related data can illuminate otherwise obscured inequities. For example, figures on COVID-19 mortality by race and ethnicity were an important tool in acknowledging and addressing the disproportionate toll of the pandemic on historically marginalized communities as well as targeting vaccine outreach efforts.

Organizations such as the Association of American Medical Colleges Center for Health Justice have also highlighted the role of data in building healthy communities and developing targeted health interventions.

“Improving community health requires accurate and relevant data to identify population health inequities, develop locally relevant interventions, and track progress toward health equity," the organization says.

A dearth of data plagues many health plans

Too many organizations still aren't accurately collecting and tracking that data, though. As recently as 2019, just, 76% of commercial health plans had incomplete race data for their members, according to a 2021 paper published in Health Affairs.

Indeed, there is an important role in collecting data that seeks to understand the impact of race, gender, sexual orientation, ethnicity, disability status on health outcomes, says the American College of Healthcare Executives.

“Getting the metrics right and having the right metrics are critical to bringing real equity to healthcare," writes Gayle Capozzalo, the Executive Director, The Equity Collaborative, in a post she co-authored for the American College of Healthcare Executives.

Keep these data guidelines in mind

The Urban Institute's Elevate Data for Equity project offers guidance on adjusting data collection to better advance equity while preventing harm to historically marginalized people. This process beings with the simple recognition that "data are not neutral" the group says: “The decisions people make about which data matter, what means and methods to use to collect them, and how to analyze and share them are important but silent factors that reflect the interests, assumptions, and biases of the people involved."

Once the data are collected and gathered, the work isn't finished. Visualization and inclusive language matters, too, notes the Urban Institute. Even the decision what group to include first in a graphic can influence a viewer's perception of who is considered the most important. The Institute also advises graphic creators to consider whether categories such as “other" have an exclusionary connotation. Colors matter and can reinforce gender or racial stereotypes.

In a similar vein, the Annie E. Casey Foundation offers guiding principles for making data analytics work. This process includes providing data transparency and evidence, empowering communities, and promoting equitable outcomes.

Ensure your digital tools advance health equity

RTI Health Advance recognizes the potential of digital tools to transform healthcare delivery while acknowledging the health equity challenges. Let us help your organization create digital health and well-being strategies with a clear focus on advancing health equity.