Healthcare stakeholders recently gathered at the "Together to Catalyze Change for Racial Equity in Clinical Algorithms" event to discuss a major hurdle for health equity: the use of race as a biological construct in research and clinical algorithms.
The event -- hosted by the Doris Duke Foundation, the Council of Medical Specialty Societies and the National Academy of Medicine -- featured presentations and panelists discussing opportunities to improve the scientific rigor of race-based considerations in medical research, tackle harmful uses of race in existing clinical algorithms and incorporate health equity considerations in clinical guidance.
Race's use in clinical algorithms
To date, race has been used in a host of clinical algorithms, typically operationalized as a biological characteristic or used as a proxy variable for patients' lived experiences, according to 2022 research published in the AMA Journal of Ethics.
In both cases, the use of race-based algorithms can perpetuate biases and cause harm, as was the case in previous iterations of the estimated glomerular filtration rate (eGFR) equation used to predict kidney disease progression.
Experts underscored that race is a social construct rather than a biological or physiological characteristic. Each individual's experience of their self-identified race can also vary significantly based on where they live, their socioeconomic status and the role racism plays in their day-to-day.
Some algorithms try to use race to represent these social determinants of health (SDOH), but the dynamic nature of SDOH can limit the utility of a race-based proxy variable. To address this limitation, work is currently underway to improve many race-based algorithms currently in use.
The link between clinical guidelines and algorithms
Part of these efforts involves changing not only the algorithms, but the clinical practice guidelines used to inform their development.
"Clinical practice guidelines -- statements that include recommendations intended to optimize patient care informed by a systematic review of the evidence -- look at an entire evidence base," to appraise the risks and benefits of potential interventions, explained panel moderator Shazia Siddique, MD, MSHP, assistant professor of medicine at the University of Pennsylvania and the associate director of scientific research at the Penn Medicine Center for Evidence-Based Practice.
These considerations include looking at how the inclusion of race and ethnicity affects the risk of bias in the studies comprising the evidence base.
"A guideline panel needs to look at the trade-off of benefits and harms in order to make a formal recommendation about clinical care [or] resource allocation and other policy decisions," Siddique continued. "Guidance documents are often an expert opinion, sometimes consensus statements, but sometimes it's just the opinion of one or two experts who might be citing papers, but those studies may be cherry-picked."
Siddique further indicated that both algorithms and guidelines can impact racial disparities, and that there is significant opportunity for bias to be introduced in the development of each. This creates additional challenges, as bias often, but not always, causes harm.
To gain a better understanding of the relationship between clinical algorithms and health disparities, Siddique and her team conducted a systematic review. The researchers found that many of these tools -- some of which include race -- actually reduce disparities.
Upon further investigation, the team found that in studies demonstrating disparity reduction, there was an awareness of preexisting disparities that the authors explicitly aimed to address. Siddique emphasized that this intentionality can make a significant difference in these studies, as the analysis showed that clinicians' implicit biases impacted real-world outcomes for patients.
Further, Siddique noted that many of the algorithms also perpetuated or exacerbated disparities, and not just in the domains of race and ethnicity. She highlighted that the use of proxy variables, in general, can cause harm to marginalized patients.
The team concluded that race and ethnicity should be avoided in prediction models, and, if used, should be justified by exploring what the variable is being used as a proxy for and why.
The findings are also relevant for clinical practice guidelines, as the use of proxy variables can highlight gaps in the evidence base that need to be addressed to improve algorithms and outcomes.
Siddique indicated that it is important to consider who is making the decisions behind potential trade-offs in outcomes, as one could choose to make a value judgment and use an algorithm, or one could choose not to use the algorithm at all and devise an alternative clinical decision-making pathway.
"It's really important, as we think about guidelines, to recognize that their effect on racial and ethnic disparities can span this exact same spectrum. Guidelines can impact disparities regardless of whether they include an algorithm," she said.
Panelists echoed this need to improve equity-driven approaches to clinical guidelines.
Race-conscious approaches to create equitable clinical guidelines
Joseph Wright, MD, MPH, FAAP, chief health equity officer and senior vice president of the American Academy of Pediatrics (AAP), argued that clinical guidance in pediatric care should take a race-conscious approach.
He emphasized the importance of medical societies' governance structures, as a specialty organization's board of directors has a critical role in shaping guidance.
It's very evident that bias is broadly embedded in practice guidance representation and in clinical algorithms, but equitable transformation within professional societies requires full leadership investment.
Joseph Wright, MD, MPH, FAAPChief health equity officer and senior vice president of the American Academy of Pediatrics
"The AAP board of directors is integrally involved in who sits on these councils and committees that write policy and drive clinical guidance. They are the ones who appoint the members. They're the ones who are accountable for making sure that the right people are in the room," Wright said, noting that these guidelines inform many facets of clinical practice, including institutional protocols and care pathways.
AAP's equity approach has been shaped by this understanding, leading the organization to take on an equity-driven initiative to critically assess over 400 of its policies and guidelines.
"The goals are to demonstrate proof of concept, to demonstrate an approach to the elimination of race-based medicine in practice guidance and policies, and then to inform how we move forward: the future direction for our society and the clinical recommendations that emerge from that work," Wright continued.
As part of this process, practice guidelines, clinical reports and technical reports will be reviewed every five years to determine whether each should be reaffirmed, revised or retired. Further, AAP will work with its sister societies -- many of which help co-author guidance -- to tackle the use of race in clinical guidelines.
"It's very evident that bias is broadly embedded in practice guidance representation and in clinical algorithms, but equitable transformation within professional societies requires full leadership investment," Wright stated.
Without stakeholder buy-in, health disparities could continue to worsen over time.
The risk of inaction on healthcare disparities
As the debate around the use of race-based clinical algorithms persists, many have raised concerns that the elimination of race as a predictor variable could lead to risk underestimations for patients most affected by conditions like heart disease.
"That's a very valid concern and something that we should address and acknowledge," said Sadiya Khan, MD, MSc, Magerstadt professor of cardiovascular epidemiology at Northwestern Feinberg School of Medicine, who asserted that stakeholders must rely on high-quality data to inform appropriate calibration and accuracy measures in clinical algorithms across all racial and ethnic groups.
"One of the most important things, as we move forward, is how do we ensure that the models are accurate and also continue to iterate on models which are, at best, going to be a probabilistic tool to help inform our practice?"
Khan noted that in the context of health equity and clinical algorithms, there are a number of steps that need to be taken to ensure that the model improves outcomes. The first is external validation, which serves to test a model's safety, reliability and generalizability.
"We need to verify these models and build trust and engagement across both scientific and public communities," she stated.
From there, Khan posited that the inclusion of individual-level SDOH and social deprivation index (SDI) measures could help reveal insights into the root causes of health disparities. Finally, she indicated that universal access to high-quality healthcare is key to improving patient outcomes.
However, Khan cautioned that individual SDOH, SDI scores and variables like zip code are still crude measures, despite their potential to help refine a patient's risk. She stated that moving forward, stakeholders will need to continue exploring the factors that influence risk and how they could be integrated into clinical algorithms.
"We have to assess [the fundamental causes of health inequity] if we want to address them," Khan explained. "So ensuring that they are included in risk assessment is just an early step, but so is ensuring that we are also thinking about how this relates to the clinical practice in terms of systematic screening, integration into our clinical care workflows, incorporation of community healthcare workers and other team members, and leveraging existing community resources and programs to really address these root causes of health inequity."
She further noted that access to high-quality, unbiased healthcare survey data can shed light on additional barriers marginalized patients may face that contribute to adverse outcomes, making such information critical in the pursuit of health equity.
In the same vein, considering the impact of racism in the context of not just clinical algorithms, but also the spectrum of healthcare delivery, is crucial.
"Fundamentally, everyone here agrees racism impacts cardiovascular disease risk and health outcomes," Khan said. "These disparities are pervasive and persistent and may actually grow and widen if we don't take action."
While acknowledging the impact of racism and investigating the causes of health inequities is key to informing future algorithm development efforts, tackling gaps in existing tools is also required.
Closing representation and validation gaps in clinical algorithms
Along with the issue of using race as a proxy variable, many clinical algorithms are trained on mostly white cohorts, necessitating the use of race-based adjustments to "correct" for risk in non-white patients.
One such tool is the Fracture Risk Assessment Tool (FRAX), which was developed to predict a patient's 10-year likelihood of hip fracture or major osteoporotic fracture. Different versions of the tool have been deployed in over 80 countries, only four of which incorporate race and ethnicity into the calculator, according to Sherri-Ann Burnett-Bowie, MD, MPH, associate professor of medicine at Harvard Medical School and clinical investigator at Massachusetts General Hospital.
"When the FRAX calculator was created for the U.S., there were ethnicity-specific adjustment factors," Burnett-Bowie explained, indicating that if a user selected a non-white race or ethnicity, the patient's data would be processed as if they had selected white and adjusted after the fact depending on the race or ethnicity they chose.
"After the output, there's this adjustment factor that occurs where you multiply that output by 0.43 [for a Black woman]," she said.
The updated output is then compared to established thresholds dictating whether a patient should be treated.
"As a Black woman, the Bone Health and Osteoporosis Foundation has set certain thresholds that say 'if you meet this 3% likelihood of having a hip fracture in the next 10 years, then you should be treated.' But if you do this calculation, and your data is multiplied by this adjustment factor, it's going to be really, really hard to meet that threshold. So this is the problem," Burnett-Bowie noted.
Burnett-Bowie further highlighted that there has been no validation of the use of race in U.S. FRAX, and that the model's race adjustment is not designed to adequately assess patients of mixed race, ethnicity or heritage.
"The adjustment makes it very hard for someone to actually meet the [treatment] criteria," she added, noting that there are tremendous racial disparities in who is receiving osteoporosis therapy, even in those who are identified as being at high risk for fracture based on FRAX.
A task force, of which Burnett-Bowie is co-chair, set out to address these issues by reviewing the use of race and ethnicity in FRAX. This systematic review was then followed up with a recommendation and research agenda for improving the tool.
The review required the task force to look throughout the literature on FRAX's use to find studies applying the tool to a diverse cohort. Of the studies assessed, only two met the criteria, one of which included a mostly white cohort.
Therefore, the review could not effectively explore how FRAX predicted fracture in Asian, Black or Hispanic individuals.
"Our task force concluded that there really is little justification for estimating fracture risk while incorporating race and ethnicity in U.S. FRAX," Burnett-Bowie said, indicating that the social construct-based nature of race and the exclusion of non-white groups by virtue of FRAX contributed to this recommendation.
Instead, the task force recommended the use of a population-based calculator to better reflect changing U.S. demographics, although such a calculator needs further investigation, Burnett-Bowie stressed. Further, the group indicated that such a calculator should also look at post-fracture outcomes, as many groups at low risk for fracture are at higher risk for disability and death.
Burnett-Bowie emphasized that internal and external efforts by healthcare organizations are key to the pursuit of anti-racism and health equity in medicine.
"Having this coalition of medical societies really encourages, if not forces, all of us to do better [in terms of racial equity in clinical algorithms]."
Shania Kennedy has been covering news related to health IT and analytics since 2022.
