Race and ethnicity are not biological but social constructs. As such, they are unreliable predictors of physiologic variation across populations. Yet, many of the tools in modern medicine still used today to diagnose, monitor, and treat disease continue to factor race and ethnicity as faulty markers of biologic variation across humans.
How can scientists and researchers create more unbiased practices and tools that engender greater equity in medicine and health? How should these clinical tools be optimized to ensure precise risk prediction based on true individual variation rather than on crude labels?
These were some of the questions tackled during the daylong conference Race and Ethnicity: Achieving the promise of precision medicine, organized by the Department of Biomedical Informatics in the Blavatnik Institute at Harvard Medical School.
The combined in-person and virtual symposium held Sept. 14 at HMS brought together thought leaders in precision medicine, genetics, bioinformatics, artificial intelligence, and computational biology to discuss promises and perils on the quest toward delivering greater equity while also realizing the benefits of precision medicine.
Over the past 20 years, insights into human biology, genomics, and computational science have brought the promise of precision medicine closer to reality, but the practice of medicine remains far from equitable.
“As scientists, as physicians, as public health researchers, we have long been aware of the racism and deep inequities in our society, inequities that affect health and disease in powerful ways,” said George Q. Daley, dean of HMS, in his prerecorded remarks. “For far too long, our society—and in many ways, the medical profession—has operated under the faulty and dangerous notion that race and ethnicity are a proxy for biology.
“The pandemic has reminded us with unsparing brutality that race is NOT a biological construct but a social one,” Daley added. “COVID-19 has exposed systemic and socioeconomic pathologies that have little to do with an individual’s ancestry and a lot to do with where and how they grew up.”
In precision medicine, promise and peril
Zak Kohane, organizer of the symposium and chair of the Department of Biomedical Informatics at HMS, opened the event with a historical perspective on the genesis of the idea of precision medicine—a notion that is hardly new.
In 1899, Henry George Plimmer, a lecturer of pathology and bacteriology at St. Mary’s Hospital in London, said to his trainees:
“You will have to acquire too, for any success to be given to you, an accurate knowledge of human nature, and you will find that it is quite as important for the doctor to know what kind of patient the disease has for host, as to know what sort of disease the patient has for a guest.”
This perspective—what kind of patient the disease has for a host—formed the basis of the precision medicine report published in 2011 that propelled a research agenda that forms the core of precision medicine.
“It asks the question: Who are you and where does who you are situate you in the intersection of environmental exposures, clinical history, family history, genomic sequencing, gene expression, and epigenetic signature, just to mention a few of the data modalities,” Kohane said. “Understanding that allows us to understand what else to expect in a given individual.”
Such knowledge informs proper risk prediction, diagnosis, prognosis, prevention, and treatment, as well as the use of individualized therapy based on individual variables, Kohane added.
“However, there is also a risk in precision medicine,” he said. “Done poorly or in ignorant ways, it can put you in a box that might only superficially pertain to you and therefore misguide the treatments you get.”
The discussion unfolded along several thematic veins:
- How precision medicine can help untangle the role of nonbiologic drivers of disease and health.
- How precision medicine can help eliminate the outdated uses of identity in clinical medicine and propel less biased, better individualized approaches that transcend crude labels.
- How the use of AI and machine-learning tools in precision medicine can carry both promise and peril.
- How the scientific and research enterprise should flip the script and engage traditionally marginalized groups not solely as advisors but as leaders in precision medicine research and interventions.
Genome versus exposome: understanding the role of environment in disease and health
An individual’s physical, cultural, and social environments play critical roles in their health and can precipitate physiologic dysfunction and lead to disease in various ways. It is a notion that every first-year medical student is introduced to on day one of their training when they learn how to take a patient history, said panelist Charlotte Owens, an obstetrics and gynecology physician and vice president and head of the Research and
Development Center for Health Equity and Patient Affairs at the biopharmaceutical company Takeda.
“We are taught to think of the home environment, education, employment, eating, activities, drugs, sexuality, suicidality, depression,” Owens said, reflecting on her training as medical student. “And this is not really just an attempt to document the story of a person’s life, but to truly understand everything that embodies their reason for being in front of you today that might have contributed to their current chief complaint.”
Yet, to date, genetics has had a disproportionately large profile in the public and professional understanding of disease risk. Scientists lack reliable ways to measure precisely how nongenetic factors can affect health and must develop better ways to measure the effects of the social, physical, and cultural environments on disease, the panelists agreed. The environmental, social, and cultural variables that modulate risk could be collectively viewed as the socio-exposome, the constellation of exposures that are nonbiological in nature but can alter biology to lead to disease.
Where people live, work, and play make critical contributions to their health, said panelist Herman Taylor, professor of medicine and director of the cardiovascular research institute at Morehouse School of Medicine, and a graduate of HMS.
He cited an illustrative example from the field of cardiovascular medicine:
“Dr. Charles Rotimi and Dr. Richard Cooper documented that although African-Americans who populate the ‘New World,’ traditionally are thought of having high blood pressure almost as a racial trait; if you look at the diaspora and get average blood pressures from West Africa, where many of us trace 80 percent of our genetic roots, to the West Indies and then on to Chicago, you’ll see a stepwise increase in blood pressure owing, at least on the surface, to the location of that set of quote ‘Black genes.’ Clearly, this is not indicative of some genetic tendency but rather of the profound impact of environment.”
Designing preventive measures and interventions—both on the clinical and policy levels—should start with defining ways to measure and study the effects of environment on human health, said panelist Chirag Patel, associate professor of biomedical informatics at HMS.
“We need to get better at defining what the environment is, measuring it, and designing our research and interventional studies to dissect the directionality between environmental exposure and health disparities and of environmental disparities to health exposure,” said Patel, who led a 2019 study that used big data and computational analysis to disentangle the roles of environment and genes in multiple diseases.
To more clearly delineate the contributions of environment and biology Patel urged precision-medicine researchers to borrow tools from colleagues in genetics.
“We can lean on our friends in the genomics world,” Patel said. “They have measurement tools and study designs that have enabled data-driven searches for genetic variants that we have not yet seen in environmental studies. I think this is where we need to go.”
On the policy level, a systems approach to human health and to greater equity would involve solving the social and environmental factors that contribute to disease development or that may prevent a patient from following a treatment plan, turning what could be a manageable condition into an unmanageable one.
“What this means in terms of precision medicine, is we have to go beyond genetics and start to tease out how we can impact the social environment around our patients,” Owens said. “Without that, if we only focus on our amazing treatment plan, our amazing surgical intervention, we will miss an opportunity to have a sustainable and even obtainable state of health for most patients.”
Toward more enlightened measurement of variation
The application of a race coefficient in formulas that measure kidney function has been one of the most widely publicized examples of the inappropriate use of race in medical care, but there are many other examples.
Such formulas and scores are “vestiges of race-based medicine that persist into our supposedly more enlightened, evidence-based twenty-first century,” Daley said. “There is an acute need to develop more refined clinical tools that ensure both performance and equity in clinical care.”
The race adjustment in assessing kidney function was a quick-fix solution to an observation that to this day remains unexplained—Black people, on average, have higher levels of creatinine, a waste substance that is removed by the kidneys. Generally, higher creatinine levels signal worse kidney function because they indicate the kidneys may be struggling to filter creatinine from the blood. In the 1990s, however, researchers observed that, on average, Black patients had higher creatinine levels, without necessarily having kidney disease, compared with other racial groups. This observation led scientists to devise a formula that accounted for this difference by incorporating a race coefficient in the calculation used to estimate a person’s kidney filtration rate. The problem, the panelists said, is that back then scientists did not bother to look beyond race—either as perceived by the physician or as described by the patient—and to question why creatinine levels may be higher in some without kidney disease.
“This is one of the lessons we should learn going forward: that you just cannot stop with race,” said panelist Nwamaka Eneanya, a health-equity researcher, attending nephrologist, and assistant professor of epidemiology and medicine at the University of Pennsylvania. “We know for a fact that there are things that affect creatinine—medications, kidney disease, muscle mass, some genetic biomarkers—none of these things were used in these original studies that were assessing a biological outcome.”
Panelists discussed several key areas that required urgent action:
- Increase transparency with patients whenever a clinician uses their identity to determine a diagnosis, prognosis, or treatment plan.
- Collect evidence and data that inform what truly drives observed differences across populations—the social, environmental, and biologic determinants thereof—and based on that knowledge develop tools and biomarkers that accurately, reliably, and consistently yield precise diagnoses, prognoses, and treatments.
- Diversify the voices in the search for solutions—with patients from underrepresented groups front and center—and bring together clinical and sociological perspectives.
- Acknowledge that biomedical science may not have all the answers and that researchers and clinicians should collaborate with other disciplines that have looked at health inequities in marginalized communities for centuries.
Panelist Arjun Manrai also cautioned colleagues to beware of false dichotomies as in the argument that race must be used in assessing kidney function to capture biological differences.
“That is a false dilemma embedded into our collective consciousness that we either use race or we are inaccurate. In reality, there is an infinite number of ways to remove race from these equations,” said Manrai, HMS assistant professor of pediatrics and of biomedical informatics at Boston Children’s Hospital, where he directs a research lab that develops machine learning methods to improve clinical decision-making tools. “It is crucial to understand the strengths and weakness of the different approaches to removing race clinical tools.”
Neil Powe, professor of medicine at the University of California-San Francisco and a graduate of HMS, cautioned that purging race from clinical algorithms alone will not solve deeper pathologies that drive racism in the practice of medicine.
“Even more concerning than writing a patient’s race in the medical record is clinicians looking at a patient and treating them on the basis of their skin color and other physical and visible characteristics—that’s what George Floyd taught us,” said Powe, who is a general internist with longstanding interest in health inequities. “It wasn’t a card with the race, it was someone making assumptions about motivations or desires or values.”
“If we want to get rid of bias—explicit or implicit—we have to go deeper in examining physicians’ biases, our own biases, not just race in clinical algorithms, that’s where our emphasis should be,” Powe said.
The role of AI and machine learning in equitable medicine
“No conversation about precision medicine would be complete without addressing what is arguably the greatest scientific disruption poised to reshape the practice of medicine: advances in computational science,” Daley said.
As recently as 10 years ago, clinical AI was more of a theoretical exploration. Today, machine-learning tools are being used in medicine to diagnose disease, make prognostic decisions, and inform treatment choice.
Panelists discussed how to ensure that clinical machine-learning tools are built and deployed equitably.
“When you hear artificial intelligence, depending on how many Westworld or Black Mirror episodes you’ve binged, you may conjure up some utopian or dystopian scenario for the world,” Manrai said. “In practice, the bias that’s baked into our medical machine-learning models is much more subtle but is just as worrisome at scale.”
Addressing the problem, the panelists said, must start with acknowledging its existence and remaining vigilant about it every step of the way—from design to testing to implementation.
“Bias is inherent in the algorithms we build because we are humans and because our human-built machines are building these algorithms,” said Ana Oromendia, senior director of product, integrated evidence at Acorn AI. “We are imperfect, we have imperfect datasets, and we have to acknowledge that there will be bias that comes with them. How we deal with this bias is ultimately what impacts patient care down the line.”
Algorithmic bias comes in two main forms—from the data on which the model is trained and the bases on which the model is built.
Panelists spoke of the need to remove the biases in the raw data on which algorithms are trained so that biases are not baked into and propagated by the model.
But just as importantly, panelists cautioned, those who design and use AI tools must be aware of and eliminate bias in modeling or in algorithmic choices that may appear neutral but may, in fact, direct the model in how and what it learns.
“With these technologies we are not aware how able they are to do as we do, not as we say, so they end up reproducing and exacerbating biases that are present in human behavior without our knowledge often,” said Marzyeh Ghassemi, assistant professor of electrical engineering and computer science at MIT. “I do think we need to have a very holistic view of all the different technical levers that we may think of are the most modern, state-of-the-art but actually don’t work well for a particular setting in health and health data.”
Another potential pitfall is so-called black-box reasoning or interpretability—the degree to which a clinician using an AI tool can understand why the model is making a given decision. In other words, the need for transparency of the “reasoning” behind the AI tool and why the tool may be choosing one diagnosis or outcome over another.
To solve some of these challenges, the panelists called for regulatory oversight on how machine learning should be tested and audited.
Another way to reduce bias, the panelists said, is to ensure diversity in the teams that build, test, and use the algorithms.
“I’ve spent the last few years studying bias in algorithms so I am deeply aware that algorithmic approaches are no panacea to inequality and, indeed, can exacerbate it if improperly applied,” said Emma Pierson, assistant professor of computer science at Cornell Tech.
In spite of this, Pierson and her fellow panelists characterized themselves as algorithmic optimists.
Pierson cited a recent example from her work as an illustration of the promise of AI in medicine. There are well-known differences in pain severity among people with knee osteoarthritis, with underrepresented groups experiencing disproportionately more severe pain. These pain gaps persist even when the severity of the condition is objectively the same based on X-rays. Pierson’s team used an AI tool to read X-ray images of knees and identified a thus-far unrecognized marker hidden in the images to explain disparities in pain scores. Because imaging assessments are often used to determine whether a person gets surgery, previously under-diagnosed patients could now qualify for treatment.
“Machine learning gives us unprecedented ability to expand the frontiers of medical knowledge, and if we choose to focus this ability on groups that we’ve previously been blind to, we have the potential to make medicine more equitable,” Pierson said.
Inclusive excellence in the science workforce as a scientific imperative for precision medicine
Gary Gibbons, director of National Heart, Lung, and Blood Institute and an HMS alumnus, said that achieving equity in health outcomes and reducing health disparities starts with achieving diversity and equity among those who are in the trenches of science and on the frontlines of medicine.
Talent is widely distributed but opportunity often is not, Gibbons said. Therefore, he added, having an inclusive and diverse workforce across the spectrum of biomedical research is fundamental to driving innovation and scientific advances.
“Someone put forward a moral-persuasive argument for diversity and inclusion, but as a group of scientists who are also data driven, I would submit that much research also suggests we would have better outcomes, better innovation, ideation, and application of technology if we have this principle of promoting inclusive excellence,” Gibbons said.
Flipping the script to expand the notion of team science
Perhaps the most critical ingredient in achieving health equity in science and medicine is engaging the very populations that equitable precision medicine is intended to benefit and serve—those who have historically been treated inequitably, said keynote speaker Tabia Akintobi, professor of community health and preventive medicine at Morehouse School of Medicine and the school’s associate dean for community engagement.
She urged the audience to consider flipping the script by engaging underrepresented communities as co-creators of science and bringing them to the table not just as advisors or a token presence but as leaders in efforts to advance precision medicine.
Akintobi defined three groups that should be a part of the cloud of thought leaders and knowledge-holders who are designing precision medicine research and interventions: First and foremost, community and patient groups, especially people with racial and ethnic backgrounds underrepresented in research, should be engaged from the inception of the study; second, those who can effect change, including policy leaders and representatives from social services and health agencies; finally, clinicians and researchers who can help define what success looks like and support the evaluation of initiatives.
Akintobi urged colleagues to think about the intentional engagement of racial and ethnic groups in the process of research—not just to provide advice and give their perspectives—but to expand notions of team science and to serve as co-creators and translators in the process.
But the true yardstick of the success of any scientific effort, she added, remains how it is applied and deployed to serve those who need it. Quoting her mentor, Ella Heard Trammell, Akintobi posed a question: “Research is good, but what do you do with the research after you get it?”
“If we are thinking about ways in which we can be more precise in the engagement of racial and ethnic groups,” Akintobi said. “We must invest in chronically being able to answer this question and ideally, for the research to be good it requires our attention, established relationship, trustworthiness with marginalized groups that we must reposition to seats of power to educate us on what works and what their ‘why’ is.”