3

Current Use of Race and Ethnicity in Biomedical Research

The previous chapter defined race, ethnicity, and biomedical research and concluded by discussing the complexity of their intersection. Building on those conceptual foundations, this chapter begins with an overview of the uses of race and ethnicity in the biomedical research process. The remainder of the chapter considers more specific applications of the use of race and ethnicity and the consequences of their use in biomedical research, clinical practice, and the development and use of medical technologies. Given that biomedical research is often conducted using existing datasets, this chapter also examines race and ethnicity in secondary data and potential biases in these datasets that can affect downstream applications, including those using artificial intelligence. Lastly, the chapter discusses the study of health disparities, which are a primary impetus for the continued use of race and ethnicity in biomedical research.

Because biomedical research is a broad space encompassing many disciplines, it is helpful to examine its operations through a general framework. The biomedical research process can be viewed as a cyclical assembly of steps (Figure 3-1). Every research study starts with a question to be answered or a problem that needs solving, from which the study originates at the conception stage. The study design stage determines the overall approach to addressing the question of interest. If the study involves human participants, the study design is followed by recruitment and data collection. Although it is common in biomedical research to concentrate on community engagement efforts as part of study design or recruitment, community engagement and partnership can occur at and benefit every stage of the research process cycle. Observational studies that make use of existing data may skip recruitment and directly assemble the study dataset after the study design. After analyzing the study data, researchers interpret the results and determine how thoroughly the research question has been answered. Results are then disseminated through journal publications, conferences, presentations, and more. The final stage in the research process cycle is an evaluation of the preceding steps to assess

lessons learned, what new questions have arisen from the study, and how to feed new information into subsequent studies.

An important additional consideration is the implementation of biomedical research findings to improve existing practice. Implementation efforts often aggregate results across multiple studies and take into account a variety of other considerations. Developing clinical practice guidelines, for example, may identify and synthesize evidence across relevant literature. In cases such as clinical trials, the creation of medical soft-

ware, and the development of medical devices, implementation will involve applying the drug, therapy, algorithm, model, or device in the health care setting. Analogously, in fields such as biosocial research, social epidemiology, and public health research, implementation may include applying an intervention that takes into account economic, behavioral, and social determinants of health. The science of implementation—details of which are beyond the scope of this report—is an evolving field that uses methods and strategies to integrate evidence-based practices, interventions, and policies into routine health care (Bauer and Kirchner, 2020). Principles of implementation science, however, may be useful in the earlier stages of biomedical and clinical research to engineer health equity (Baumann and Cabassa, 2020; Reese et al., 2024).

FUNCTIONS OF RACE AND ETHNICITY IN BIOMEDICAL RESEARCH

Considering the range of studies that fall under the heading of biomedical research, it can be helpful to group kinds of biomedical research into broad categories to assess how race and ethnicity are included in those studies. Table 3-1 defines research strategies according to translational stage. Depending on the question being asked and on the relevant sub-discipline(s), biomedical research leverages data from various modalities throughout these translational stages. These modalities include human molecular data, clinical indicators, electronic health records (EHR), claims and billing data, public health data, and more. Given the many modalities for conducting biomedical research and the many reasons that race and ethnicity data may be collected and used, there is significant variation in how the information is recorded, if it is recorded at all.

This section describes different types of biomedical research, how race and ethnicity are often used currently, and the relevance of race and ethnicity to each type of research study. Of note, race is often used as a proxy for other concepts that might more specifically address the research question of interest, so what race attempts to represent in different types of biomedical research is highly context dependent (see Chapter 5).

Basic science (T0) uses laboratory-based techniques, such as preclinical cell and animal models, to investigate biological mechanisms. In general, race and ethnicity are less relevant for characterizing most fundamental biological phenomena, such as in developmental biology, because these biological mechanisms (e.g., DNA replication,

TABLE 3-1 Types of Biomedical Research Study

SOURCE: Adapted from Blumberg (2012), https://www.nature.com/articles/nm.2632, reprinted with permission from Springer Nature.

mitosis) are shared among humans and often across species. Currently, in this type of basic science research, race and ethnicity are commonly not considered, and the constructs have lower relevance to research focused only on understanding fundamental biology (see further discussion in Chapter 6, section “Basic Science and Early-Stage Biomedical Research”).

Translational research (T1–T4) uses basic science to inform biomedical innovation, with the goal of improving wellness, health, and health care. Early-stage translational research (T1) focuses on developing proof of concept or determining the efficacy and utility of an intervention, such as a drug that has been shown previously to have desirable biological effects in cultured cells or in an animal model. These studies are performed under ideal or highly controlled circumstances. T1 research includes small Phase I clinical trials that test a specific intervention; however, not all T1 research is done with a clinical trial. This translational stage also encompasses exploratory research with human participants or human-derived samples to evaluate new assays, characterize a condition, or validate a hypothesis; these studies do not affect treatment decisions for patients, so they are not clinical trials (NIH, 2017). Race and ethnicity are sometimes used in such studies to ensure a diverse sample of participants is recruited, and clinical trials are likely to have strict reporting requirements. Even so, researchers may, though not always, predicate these research efforts on biomedical circumstances that would not vary by race. Thus, the relevance of race and ethnicity to T1 research depends on the specific research questions and ranges from low to moderate.

Interventions that advance to the next research stage (T2) then receive subsequent evaluation for effectiveness, suitability, and utility in “real world” circumstances involving different people with varied history, contexts, resources, values, and preferences. In this type of research, collecting race and ethnicity data is typically required for recruiting diverse sample populations. Sometimes these constructs are also used as imperfect proxies to understand how implementation, resources, values, and preferences affect intervention effectiveness. As the social context becomes increasingly relevant to these research questions, the relevance of race and ethnicity to the research context increases as well.

Research at stages T3 and T4 may also highlight considerations of comparative effectiveness. The impact and import of a given treatment potentially vary between and within race and ethnicity groups. Consequently, it is important to consider variation regarding social factors within and between race and ethnic groups that might affect outcomes. As research is translated to patients and clinical practice, patients’ and clinicians’ values and preferences may differ between and within race and ethnic groups. Consequently, collecting data on race and ethnicity can clarify treatment effects. Even though race and ethnicity may not themselves be the underlying mechanism for differences (see Chapter 5), there can be a role for collecting the data, as is often required. Collecting race and ethnicity information may be useful to clarify whether diffusion of innovation varies by the race and ethnicity of care teams and patients. For example, this information could be used to study whether the innovations reach racially or ethnically diverse care teams, whether these teams adopt these innovations, and whether racially and ethnically homogeneous care teams implement care innovations in racially and ethnically diverse patient groups. These types of inquiry help ensure that all groups

are positioned to benefit equally from treatment innovations. Therefore, in these lines of research race and ethnicity can be highly relevant.

The social constructs of race and ethnicity can serve multiple functions in biomedical research, depending on the category of research and the goals of the particular study. In addition, within a single study race and ethnicity can serve various purposes across different stages of the research process. For instance, racial and ethnic categories may be used in formulating the research question and study design, in recruitment to ensure a diverse sample of research participants, during data analysis to stratify data and evaluate interactions, and then to make inferences based on the analysis. In addition to these general functions, the social categories of race and ethnicity have sometimes been used as an input variable in the subset of biomedical research that is used to construct clinical algorithms and decision-making tools (see subsequent sections in this chapter for more information about race correction).

Research That Uses Results from Translational Biomedical Research

Beyond the translational research spectrum, there are other research domains that draw from the findings of biomedical research. Though a detailed examination was outside the scope of this study, it is important to recognize the impact that biomedical research has in these areas. According to a 2009 Institute of Medicine report, “comparative effectiveness research is the generation and synthesis of evidence that compares the benefits and harms of alternative methods to prevent, diagnose, treat, and monitor a clinical condition or to improve the delivery of care. The purpose of comparative effectiveness research is to assist consumers, clinicians, purchasers, and policy makers to make informed decisions that will improve health care at both the individual and population levels” (IOM, 2009). Comparative effectiveness research typically prioritizes patients’ values and preferences in decision making. Race and ethnicity are sometimes treated as imperfect proxies for patient awareness, resources, values, or preferences related to decisions. In current practice, race and ethnicity may also influence whether or how clinicians or organizations provide or frame treatment options. Consequently, careful consideration of issues related to race and ethnicity is highly relevant in this research domain.

Implementation science is the study of the conditions facilitating the systematic uptake of proven effective interventions. This research domain helps translate biomedical research into effective practice under real-world conditions. Race and ethnicity, along with a range of other demographic information, could be relevant in identifying a diverse sample of patients and communities to determine how best to deliver an intervention to meet their needs (Bodison et al., 2015; Mensah, 2019).

Quality improvement is a means and method to systematically monitor and measure how to augment the performance of evidence-based practice. Quality improvement focuses on whether a health care facility executes effective care in a patient-centered manner and seeks strategies to deliver that patient-centered care more efficiently. Examining whether and how interventions reach all patient groups requires that race and ethnicity data be collected.

HISTORY AND CURRENT PRACTICES OF RACE CORRECTION

Race correction, also known as race norming or race adjustment, is the practice of developing clinical algorithms or practice guidelines that adjust their outputs based on a patient’s race or ethnicity (Vyas et al., 2020). Physicians use these tools and guidelines, which often collect information related to symptoms, medical and family history, and other personal details along with race and ethnicity, to inform their clinical decisions and make assessments regarding an individual’s risk of developing certain medical conditions (Doris Duke Foundation, 2023; Vyas et al., 2020). In addition to clinical algorithms and guidelines, adjustments based on race and ethnicity can also be hardwired into medical equipment (Lujan and DiCarlo, 2018).

The notion that a person’s racial or ethnic identity determines their susceptibility to disease has been attributed to a historic belief that there are inherent biological differences between members of different racial and ethnic groups. Thomas Jefferson, for example, noted “a difference of structure in the pulmonary apparatus” between enslaved individuals and European colonists, while Samuel Cartwright, a prominent physician of the time, argued that there was a 20 percent deficiency in the pulmonary function of Black people (Braun, 2014; Hammond and Herzig, 2009; Lujan and DiCarlo, 2018). Thus, slavery and forced labor were perpetuated under the assumption that they would help revitalize the health of Black Americans who were allegedly more prone to disease than their White counterparts (Lujan and DiCarlo, 2018). The practice of adjusting or correcting for an individual’s racial or ethnic identity became ingrained throughout various subfields in medicine, ostensibly to achieve greater precision in predicting a patient’s risk level. Despite recent changes (e.g., the development of a race-free calculator to assess kidney function), race correction persists in medicine. It is not a simple matter, however, to remove race or ethnicity from clinical tools, algorithms, and guidelines. How to properly address the relationship between race and ethnicity and health outcomes without exacerbating existing health disparities is an area of active research (see Chapter 5 for a reconceptualization of the relevant variables and Chapter 6 for this committee’s recommendations). Table 3-2 lists examples of race correction that were or are currently used in medicine.

Examples of Race Correction in Clinical Practice

Pulmonology

Spirometers are medical devices used to diagnose respiratory disease by measuring the volume of air let out after a deep breath (Anderson et al., 2021; Braun, 2014). Most of these devices apply a correction factor of 4–6 percent smaller lung capacity for Asians and 10–15 percent smaller lung capacity for Black individuals (Anderson et al., 2021). These corrections are hardwired into the software of spirometers and are automatically applied to their outputs, often without much awareness or consideration from the physicians using these instruments (Anderson et al., 2021; Lujan and DiCarlo, 2018; Wright et al., 2022). There is no standardized correction factor for mixed-race individuals (Anderson et al., 2021). Furthermore, this example of race correction perpetuates the

TABLE 3-2 Examples of Clinical Calculators and Tools That Incorporate Race Correction

SOURCES: Clinicalalgorithmswithrace.org (2023); Visweswaran et al. (2023) CC BY 4.0.

notion of innate biological differences, or race science, that has historically been used to rationalize the oppression of racial and ethnic minority groups. The European Respiratory Society and the American Thoracic Society have stated that adjustment factors for race are not appropriate and are discouraged (Diao et al., 2024; Stanojevic et al., 2022). The Global Lung Function Initiative in 2022 replaced race-based equations with

new equations that do not incorporate race but instead use a weighted average across race groups. Comparing the previous race-adjusted equations with the race-neutral equations, a recent study reported that both classes of equations provided similarly accurate predictions of respiratory outcomes (e.g., respiratory symptoms, new-onset disease, and death from respiratory causes) but assigned different disease classifications, occupational eligibility, and disability compensation for millions of people (Diao et al., 2024). Despite similar performance in accuracy, these revised classifications had broad clinical, occupational, and financial implications that differed across racial and ethnic groups (Diao et al., 2024). This may be due to small differences related to lung-function thresholds or differences in sensitivity and specificity that are not accounted for when examining accuracy overall, emphasizing the importance of comprehensively evaluating potential tradeoffs (Diao et al., 2024).

Nephrology

Black Americans are almost four times more likely to develop end-stage kidney disease than White Americans (NIDDK, 2023), and they face other disparities in kidney disease progression (Ahmed et al., 2021). The contributing factors are not completely understood, but social determinants of health (Nicholas et al., 2015; Norton et al., 2016; Powe, 2021) and a higher prevalence of genetic variants that are associated with increased risk for kidney disease (Drawz and Sedor, 2011; Freedman et al., 2018; Friedman and Pollak, 2020) likely play a role in this disparate burden of disease. Powered by the social justice movement following the murder of George Floyd and other young Black people, medical communities responded with a heightened awareness of health and health care inequities (Powe, 2022), and trainee contributions were instrumental to galvanizing change in nephrology (Heffron et al., 2022). One area of focus became equations for estimating glomerular filtration rate (GFR), a key measure of kidney function, and specifically the use of race as a categorical variable in the equation.

In 1998, Camille Jones and colleagues at the U.S. National Institutes of Health (NIH) used the National Health and Nutrition Examination Survey to show that Black men and women at all ages had higher serum creatine levels than their White counterparts and argued that there should not be one level of normality until this finding was clarified (C. Jones et al., 1998). With a gold standard of directly measured (rather than estimated) GFR, Levey et al. found a correlation of higher serum creatinine levels among Black people, and this finding was replicated in both U.S. and European datasets (Inker et al., 2021; Pottel et al., 2023). These observations led to the incorporation of race into the Modification of Diet in Renal Disease (MDRD) estimated GFR (eGFR) equation in 1999 (Levey et al., 1999) and subsequently in 2009 in the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) eGFR equation (Stevens et al., 2008). Beginning in 2017, the Beth Israel Deaconess Medical Center in Boston officially removed the race coefficient and provided a range of estimated values with and without the correction factor because of concerns that the eGFR overestimated the kidney function of Black individuals and might deprive Black persons with chronic kidney disease of the opportunity to obtain a timely kidney transplant (Hoenig et al., 2022).

Several other medical centers began to do the same, by removing the race coefficient from the calculation without fully characterizing potential effects (Powe, 2020). The National Kidney Foundation and American Society of Nephrology formed a task force in 2020 to reassess the use of race in the calculation of eGFR. The task force used a rigorous process to recommend a newly developed CKD-EPI race-free equation using serum creatinine as well as greater use of cystatin C, a biomarker for kidney function that does not vary by race (Delgado et al., 2022). The new equation changing eGFR for all patients does not include race in the calculation and reporting, and special consideration was taken to ensure that the potential consequences do not disproportionately affect any one group of individuals. This equation was rapidly adopted by medical centers and laboratories across the United States as well as by the Organ Procurement Transplant Network (Genzen et al., 2023).

Obstetrics

In the field of obstetrics, an individual’s race and ethnicity can influence their risk assessment for certain conditions and what treatment or procedures are recommended. For example, the Institute of Medicine recommended different diagnostic criteria based on race for detecting anemia among pregnant women (IOM, 1993). Until recently, this guideline was supported by the American College of Obstetricians and Gynecologists (Brown et al., 2022). Moreover, the commonly used vaginal birth after Cesarean (VBAC) algorithm applies two different correction factors for Black and Hispanic individuals in order to calculate the likelihood of successfully giving birth vaginally after undergoing a Cesarean section in a previous pregnancy (O’Brien and Clare, 2023). The correction factors assigned to patients from either of these backgrounds result in outputs that predict a lower likelihood of a successful VBAC, meaning that these patients are systematically less likely than White patients to attempt a vaginal delivery, which is typically less risky than a Cesarean delivery (O’Brien and Clare, 2023; Wright et al., 2022). Although other social factors, including insurance type and marital status, were also associated with the likelihood of successful VBAC, these were not included in the 2007 calculator (Vyas et al., 2019). Moreover, these correction factors can be traced back to a long history of race science and descriptions of anatomical differences that deemed the pelvises of White women more suited to childbirth than those of Black and Hispanic women (Vyas et al., 2019). More recently, a race-free VBAC calculator was developed (Grobman et al., 2021) and was shown to be accurate in a diverse cohort of patients at an urban medical center (Adjei et al., 2023). However, knowledge of the new race-free calculator is not widespread (Cron et al., 2024), and it will take a concerted effort to challenge misplaced beliefs about the role of race and ethnicity in this area of medicine.

Cardiology

The American Heart Association Get with the Guidelines–Heart Failure Risk Score predicts the risk of death for patients admitted to the hospital for a heart condition (Peterson et al., 2010; Vyas et al., 2020). This score assigns three additional points to

non-Black individuals, while Black patients are automatically categorized as lower risk. The Society of Thoracic Surgeons’ risk calculator estimates the likelihood of experiencing complications or death during heart surgery and includes information about patients’ race and ethnicity because of differences in surgical outcomes between different groups (Shahian et al., 2018; Vyas et al., 2020). Using this calculator, the risk of death for a White patient undergoing an isolated coronary artery bypass is 0.492 percent but jumps by almost 20 percent if the patient’s race was changed to “Black/African American” (Shahian et al., 2018; Vyas et al., 2020), thereby reducing the likelihood that this procedure would be recommended for Black patients. A new sex-specific, race-free equation PREVENT that predicts risk of cardiovascular events was recently introduced (Khan, 2023).

Hematology

Approximately two out of three people with African ancestry in the United States have the Duffy null phenotype, a non-expression of the Duffy antigen on red blood cells, which can manifest as chronic neutropenia in otherwise healthy individuals (Merz et al., 2023). This condition is known as benign ethnic neutropenia, but individuals with this phenotype are often mislabeled based on their race or ethnicity as suffering from neutropenia, a condition that involves a deficiency in a specific type of white blood cell, which increases susceptibility to infection and can indicate underlying bone marrow dysfunction (American Cancer Society, 2023; Atallah-Yunes et al., 2019; Doris Duke Foundation, 2023). Patients with benign ethnic neutropenia are not at increased risk of infection (Atallah-Yunes et al., 2019), but misdiagnosis may subject them to unnecessary medical testing, exclude them from clinical trials, and prevent them from receiving appropriate chemotherapeutic drugs (CMSS, 2023; Merz et al., 2023).

Pharmacogenomics

Pharmacogenomics is a specialty that examines how a person’s genetic variants influence their response to a drug for the purpose of choosing a more individualized therapy (Goodman and Brett, 2021). Racial and ethnic categories are often used in this field to stratify genetic risk based on the assumption that these categories can adequately identify populations that have a high or low prevalence of specific genes, thereby enabling physicians to refer high-prevalence groups for additional testing and care (Goodman and Brett, 2021). However, pharmacogenetic screening based on race or ethnicity can be limiting or even misleading. For example, the American College of Rheumatology recommended that individuals who identify as Southeast Asian or African American test for the HLA-B*5801 allele before taking allopurinol, a medication for gout, because of studies demonstrating a higher prevalence of this allele in these groups (FitzGerald et al., 2020; Goodman and Brett, 2021). The presence of the HLA-B*5801 allele has been associated with allopurinol-induced severe cutaneous adverse reactions, which can be deadly. The limitations of this guideline, however, are exemplified by the vast amount of genetic variation present within certain racial,

ethnic, or geographic populations, which can exceed the variation across these groups. Data from Switzerland attest to this fact; despite Switzerland’s smaller size and lower racial and ethnic diversity compared to the United States, HLA-B*5801 frequencies vary considerably throughout the country, making race and ethnicity poor proxies for capturing this variability (Goodman and Brett, 2021).

U.S. Food and Drug Administration (FDA) product labeling contains prescribing information for certain drugs that includes race-adjusted indication, dose, and monitoring (Clinical Algorithms with Race, 2023; Visweswaran et al., 2023). For example, initial dosage for warfarin¹ is influenced by several factors including age, race, genetics, and body weight, among others, and dosage for omeprazole² is indicated by race because of a correlation to CYP2C19 genotype. Some of the justification for using race in the context of prescribing medication is based on pharmacokinetics or pharmacogenetics studies where associations were made with racial groups. However, because genetic variation within racial and ethnic groups can be greater than the variation across such groups, race or ethnicity-based pharmacogenetic decision making can be limited by intrapopulation genetic variation along with the sociopolitical nature of race and ethnicity categories themselves (Goodman and Brett, 2021; O’Brien et al., 2021). As discussed in Chapter 5, though race is sometimes used to assess patients’ disease risk, race is not a substitute for unmeasured biological indicators of disease risk and, in the case of pharmacogenomics, relying upon race can be limiting for assessing dosage or adverse drug events.

Consequences of Race Correction and Considerations for Future Practice

As the examples in the previous section make evident, there are many potential consequences of applying race correction in clinical algorithms and decision-making tools (Siddique et al., 2024). The evidence base that resulted in race correction or adjustment in clinical algorithms and decision-making tools is based in part on beliefs that racial and ethnic groups are distinct and are biologically different. Algorithms that assess patients’ risk for developing certain conditions and guide treatment protocols can direct attention or resources away from racial or ethnic minority groups (Vyas et al., 2020). Ahmed et al. (2021) explored how the race correction factor influenced the diagnosis and care of chronic kidney disease and found that over a third of Black patients who were part of the study would have hypothetically been reclassified to a more severe stage of the disease had race been removed from the CKD-EPI calculator. This algorithmic bias holds implications for Black patients’ ability to enroll in care coordination programs, thereby limiting the resources available to them and potentially exacerbating their disease burden. The reliance of these clinical tools on variable and subjective categorizations such as race and ethnicity can also obfuscate the role of social determinants of health (Lujan and DiCarlo, 2018). Factors like the

___________________

¹ https://www.accessdata.fda.gov/drugsatfda_docs/label/2011/009218s107lbl.pdf (accessed October 16, 2024).

² https://www.accessdata.fda.gov/drugsatfda_docs/label/2016/019810s102,022056s019lbl.pdf (accessed October 16, 2024).

high prevalence of food deserts in low-income neighborhoods, exposure to toxins in the environment, high rates of incarceration, and the physical and mental stress of experiencing racial discrimination may be far more consequential to health outcomes and are likely inadequately captured by simply checking “Black” or “White” in a clinical risk prediction tool (Lujan and DiCarlo, 2018). Research is underway to identify the impact of social factors on disease rather than using race as a proxy for factors that can be directly measured.

Adjustments based on race and ethnicity within clinical diagnostic tools can also lead to misdiagnosis or delayed diagnosis of certain conditions. The treatment of COVID-19, for example, has been complicated by the growing prevalence of restrictive ventilatory dysfunction in patients, a condition that is detected through spirometry (Anderson et al., 2021). Physicians could miss this diagnosis if they have become accustomed to associating certain racial or ethnic groups with having a lower lung capacity at baseline compared with other groups. Furthermore, these groups can be further disadvantaged in the health care system by having a higher likelihood of being recommended riskier or more invasive medical procedures. The application of the VBAC calculator illustrates this point; despite VBAC being associated with a range of positive maternal health outcomes, such as decreased morbidity and a lower risk of future complications, Black women are less likely to be recommended for VBAC and to experience these benefits on account of their racial identity, exacerbating existing disparities in maternal morbidity and mortality (O’Brien and Clare, 2023). Thus, the rationale behind including a patent’s racial or ethnic identity as a means of individualizing medical intervention does not always translate to how these tools are implemented in everyday clinical practice and can result in more harm than good in some cases.

Given the controversies surrounding the use of race correction in clinical algorithms and tools, there have been efforts in some scientific and medical disciplines to discontinue this practice. The NYC Coalition to End Racism in Clinical Algorithms (CERCA), for example, has encouraged health systems, hospitals, medical schools, and clinicians in private practice to retire the use of race correction in at least one algorithm at their facilities within 2 years (O’Brien and Clare, 2023). In addition, a joint task force established by the National Kidney Foundation and the American Society of Nephrology decided that race correction should no longer be applied to GFR calculations (Delgado et al., 2022). Health systems across multiple universities, including the University of Washington, the University of California, San Francisco, the Beth Israel Deaconess, and Vanderbilt University Medical Center, removed race from their eGFR calculator (Cerdeña et al., 2020).

Preventing the harmful consequences of race correction on health outcomes is not as simple, however, as removing the race variable from clinical algorithms; each tool and context is different and needs comprehensive examination. There are tradeoffs because race correction can have beneficial effects as well, such as when race is intentionally included to counteract a known disparity that may be tied to bias (X. Zhang et al., 2018). A recent report from the Agency for Healthcare Research and Quality found heterogeneous effects of clinical algorithms on health disparities,

regardless of whether the algorithms explicitly include race or ethnicity as an input variable; the report identified 5 algorithms that may reduce racial and ethnic disparities, 13 algorithms that may worsen disparities, and 1 with no effect on disparities (Siddique et al., 2024). Results can be complicated even for a single clinical tool—while removing race from eGFR calculations will, for Black patients, increase their addition to the kidney transplant list and the likelihood of being diagnosed with chronic and severe kidney disease, it may reduce their access to other treatments such as chemotherapy and potentially decrease their eligibility for clinical trials (Tipton et al., 2023). Moreover, how algorithms are implemented in the real world can also impact disparities (Siddique et al., 2024). These tradeoffs may also vary across different racial and ethnic minority groups, as demonstrated in a prediction model for lung cancer screening eligibility that tested the potential effects of removing race from the calculation and found that disparities might improve for Hispanic and Asian groups but could worsen for Black populations (Landy et al., 2023). (See Chapter 4 for a review of existing guidance on race and ethnicity in clinical practice guidelines, clinical algorithms, and clinical AI algorithms.) This section underscores how some current clinical practices can be traced back to a history of race science. The examples of race correction discussed here refer primarily to mathematical correction factors (e.g., applying a percentage decrease to a result) or are based on traditional statistical methods and calculations (e.g., regression analyses) (see Igo, 2007 and Zuberi, 2003 for background on race in statistics). Looking toward the future, artificial intelligence (AI) will undoubtedly reshape how clinical algorithms are both developed and implemented. The rapidly expanding use of AI in research and health care is bringing even greater awareness to potential bias in datasets and clinical algorithms (see subsequent section “Race and Ethnicity in Clinical AI Algorithms” for further discussion).

RACE AND ETHNICITY IN MEDICAL DEVICES

Skin color and pigmentation are relevant to the design and function of some medical devices that use optical technology. Failing to test these devices with a broadly representative population can result in data that are not equally reliable across the full spectrum of skin tones, contributing to differential performance of these technologies and health disparities, which may reflect larger inequities in health outcomes. When optical devices do not work accurately for people across the full range of human skin tones, this is at times referred to as “racial bias” in the design and engineering processes. However, it is important to note that skin color and race should not be conflated in seeking to improve these technologies in the future. Medical devices that produce racially biased results include pulse oximeters, transcutaneous bilirubinometers, and photoplethysmographic sensors, which use optical technology to measure blood oxygen levels, serum bilirubin levels, and heart rate and rhythm monitoring, respectively, as well as forehead thermometers, which use infrared technology to measure body temperature. Examples of the biased outputs from these devices are shown in Table 3-3.

TABLE 3-3 Examples of Medical Devices with Differential Performance Across Racial and Ethnic Groups

SOURCES: Adapted from clinicalalgorithmswithrace.org (2023). Content from Bhavani et al. (2022); Holder and Wong (2022); Merid and Volpe (2023); Varughese et al. (2018).

Pulse Oximetry: A Case Study of Race and Ethnicity in Biomedical Research and Medicine

As a window into larger questions of why diverse representation remains important to consider in biomedical research and medical device design, the pulse oximeter offers a useful case study. In recent years, the pulse oximeter has become commonly pointed to as “poster child” (McFarling, 2022) for the need for more diversity in science and medicine, including during the design and engineering phases of health technologies as well as during safety assessments. Pulse oximeters estimate blood oxygen saturation via color sensing, assessing the shade of iron-containing hemoglobin—which is a cooler color when desaturated with oxygen and a warmer red color when fully saturated.³ Because the light emitted from “wearable” devices passes through not only blood but also surrounding skin and tissues, optical readings can be affected by melanin and other chromophores in the skin. This is why diversity in safety testing groups and calibration

___________________

³ https://www.howequipmentworks.com/pulse_oximeter/ (accessed October 16, 2024).

matters so much, yet it historically has often been inadequate—with important exceptions. Decades ago, physicians at the University of California, San Francisco Hypoxia Lab realized they were testing a color-sensing device on mostly White populations and noted many years’ worth of overlooked reports of unequal errors. Their follow-up study showed that pulse oximeter devices in hospitals at the time of their research indeed did not meet FDA safety thresholds for people of color (Bickler et al., 2005; Feiner et al., 2007). At the time, there was little public response.

During the early COVID-19 pandemic, optical devices were increasingly emphasized in homes as well as clinics, positioned as “biomarkers” to be used in triage, and played a role as gatekeepers to resources such as admissions to emergency departments and supplemental oxygen. Putting this reality together with the history of concerns about such devices, a network of social scientists again brought this issue to the attention of physicians (Benjamin, 2019; Moran-Thomas, 2020; Valley, 2023), including a team that examined the issue for the first time using meta-data across major hospitals. Published in the New England Journal of Medicine, this team’s study showed that errors were three times as likely for Black hospitalized patients as for White patients (Sjoding et al., 2020). A wave of follow-up studies found that a disproportionate number of Black hospitalized patients experienced “occult hypoxemia” (Valbuena et al., 2022)—a new designation created to capture the clinical significance of seemingly small discrepancies in device accuracy (Sjoding et al., 2023). Other researchers found that these errors correlated with Black hospitalized patients receiving less oxygen supplementation in intensive care units (Gottlieb et al., 2022). Further studies revealed that disparate device discrepancies were associated not only with delayed treatment but also with sequelae such as higher rates of consequent organ dysfunction and mortality for Black hospitalized patients (Wong et al., 2021). Notably, while the categories “Black” and “White” were used as proxies at the time that these hospital data were collected, other studies have shown such inaccuracies have implications across the spectrum of skin tone variability (Feiner et al., 2007) and will require ongoing attention and nuance around self-identification and skin tone in the future. Oximeter inaccuracies—leading to “hidden hypoxia” being three times more common among Black patients than White patients (Sjoding et al., 2020)—were also found among hospitalized children (Andrist et al., 2022) and, more subtly, among preterm infants (Vesoulis et al., 2022).

Problems with Other Optical Sensing Devices

As noted in Table 3-3, inaccuracies have also been identified in other optical sensing devices, including infrared thermometers (Bhavani et al., 2022), heart monitors (Bent et al., 2020), and bilirubinometers (Varughese et al., 2018). An expanded crowd-sourced list is currently underway.⁴ There is growing concern that errors in devices’ direct inputs into algorithms can interact with—and cause unequal errors—in the artificial intelligence (AI)-mediated algorithms now becoming increasingly prominent across hospital

___________________

⁴ https://clinical-algorithms-with-race.org/devices (accessed October 16, 2024).

computing systems (Zou and Schiebinger, 2021). As part of broad movements toward decentralized care, there is increasing emphasis on at-home wearable device measurements, often without sufficient attention to equity and accuracy for all device wearers. There is also a general push toward “surrogate measures” that substitute long-standard lab readings (e.g., highly accurate blood tests) for wearable readings from devices such as oximeters (Carvalho et al., 2022). This shift is occurring in a range of algorithms used in triage, which worryingly can incorporate measurements known to be racially biased into the computing used across health care systems. Engineers are now working to create novel solutions to these issues and to imagine more equitable optical devices (Jakachira et al., 2022; McFarling, 2022), but sustained attention is needed across many sectors in order for more equitable designs to be realized at market scale. With unequal optical device hardware changing slowly and the computing and AI systems that use their inputs changing quickly, present systems are not yet capable of ensuring equal safety.

Lessons Learned

These examples illustrate what can be missed by failing to consider racial and ethnic diversity in medical device design and testing. Overlooking characteristics like skin tone can lead to discrepancies in device performance and health outcomes. Caution is warranted, though, before simply eliminating the consideration of race, which can be used to address bias and inequities. Rather, characteristics like skin pigmentation, which do not universally correlate to specific racial or ethnic categories, must be taken into consideration. It is important to learn from history while trying to correct present errors. Pulse oximeter performance, for instance, cannot be easily fixed by a race-based correction factor because issues of skin color and race are distinct (Patwari et al., 2024). Some devices may need to be redesigned, requiring a greater investment than a quick fix might entail. More fully understanding device inaccuracies (Okunlola et al., 2022) can open the doors to opportunities for much-needed future innovations.

RACE AND ETHNICITY IN SECONDARY DATA ANALYSIS

Many biomedical studies rely on preexisting datasets to address specific research questions. These previously collected datasets, or secondary data, can be reused or reanalyzed in the service of new research questions in biomedical research. Broadly, secondary data used in biomedical research can be legacy data, administrative claims or financial health care records, and EHRs. Race and ethnicity can be captured in secondary data in myriad ways, including through self-identification, via census or administrative data, from surrogate markers, from clinician and researcher determinations, or by a combination of any of these methods. Each type of secondary data has unique considerations for use as well as notable limitations.

Legacy data is made up of research data derived from past studies, such as clinical trials, genetic and genomic studies, and epidemiological studies. These datasets typically remain static after the completion of a given study. While having access to a wealth of data from previous research can help spur the development of new studies and generate

new insights on a similar topic, there are major challenges that accompany the reuse of race and ethnicity research data. One limitation is the longstanding history of underrepresentation of racial and ethnic minority groups in clinical trial and genomic datasets. For example, an analysis of genomics studies through the middle of 2021 found that 86.3 percent of participants were of European descent, 5.9 percent were East Asian, 1.1 percent African, 0.8 percent South Asian, and 0.08 percent Hispanic/Latino (Fatumo et al., 2022). A review of 20,692 U.S.-based studies in ClinicalTrials.gov from 2000 to 2020 demonstrated a similar pattern, with White individuals making up the majority of the 4.76 million trial enrollees (Turner et al., 2022).

In addition to the lack of representation of minority populations and the resulting limited generalizability of study findings, many clinical trial datasets are missing race and ethnicity data altogether. For example, only 43 percent of the 20,692 clinical trials reviewed by Turner et al. (2022) reported any race or ethnicity data. Furthermore, the methods used to collect the race and ethnicity data may be ambiguous or questionable, particularly for datasets that are decades old. These datasets are prone to aggregating data into categories like “Other” but with little or no explicit rationale. Thus, researchers who want to use legacy datasets have to contend with race and ethnicity data that are potentially inaccurate, incomplete, or of problematic origin.

Administrative claims data, or data collected from health care transactions, are generated from billing and payment records related to medical services, including insurance claims submitted to Medicare, Medicaid, and private insurance companies. As with legacy data, administrative claims data contain incomplete race and ethnicity information, due to nonreporting or misreporting. In addition, the data may be further complicated because the racial and ethnic categories of some people in the datasets change over time. This latter problem is seen in the Medicare enrollment database, which obtains race and ethnicity data from the Social Security Administration. When individuals were applying for social security numbers from 1935 to 1980, the only racial categories available were White, Black, or Other, whereas in 2024, there are seven racial and/or ethnic categories (Nead et al., 2022). There is also a lack of standardization in how states and hospitals collect this information. Medicaid claims data, for example, are collected on a state-by-state basis, with most states opting to collect self-reported race and ethnicity, yet often filling in missing or incomplete data based on a person’s name, language, and geographic location or by matching with data from other states (Nead et al., 2022). Given these limitations in race and ethnicity data within administrative claims data, researchers intending to use this type of data will likely need to navigate the vast amount of geographic and temporal variability in existing race and ethnicity datasets.

EHR data consist of large datasets collected from multiple hospitals. The use of race and ethnicity information from EHR data poses similar challenges to those found in other secondary data types, because the data often vary across hospitals and institutions and are often incomplete. EHR systems are widely used in the United States to capture data on clinical encounters. As of 2021, 96 percent of all non-federal acute care hospitals and nearly four in five office-based physicians had implemented a certified EHR system (Office of the National Coordinator for Health Information Technology, 2023). Originally, the primary purpose of EHRs in the United States was to support

clinical care, financial billing, and insurance claims. Data collected in EHRs are now extensively used for secondary purposes such as clinical research, improving health care practice, and large-scale analyses for the creation and validation of predictive algorithms. Federal agencies such as the FDA have been empowered by legislation such as the 21st Century Cures Act to increase the use of real-world evidence based on real world data, such as EHR and administrative claims data. Statistical and AI methods are increasingly applied to EHR data to study patient cohorts for various clinical and research applications, such as phenotype extraction, precision medicine, intervention evaluation, and disease prediction, detection, and progression. Examples of federally funded, privately funded, and community-based initiatives that are collecting EHRs of millions of individuals include the Patient-Centered Outcomes Research Institute’s PCORnet,⁵ the National COVID Cohort Collaborative,⁶ TriNetX,⁷ Epic Cosmos,⁸ NIH All of Us Research Program,⁹ VA Million Veteran Program,¹⁰ the Observational Medical Outcomes Partnership,¹¹ the NIH-funded ENACT network,¹² the Consortium for Clinical Characterization of COVID-19 by EHR,¹³ and the University of California Health Data Warehouse.¹⁴

EHRs contain a wide range of data types that characterize the health conditions of individuals. These data types include demographics, vital signs, medications, diagnoses, procedures, laboratory test results, clinical imaging results, and clinical notes. Demographic data often include race and ethnicity, which are ascertained in various ways: patients complete forms as part of the registration process, which are then transcribed into the EHR by a registration clerk, or registration clerks enter responses after asking patients about their racial and ethnic identification. Given the different methods that can be employed when collecting this information, the extent to which data are missing or erroneous can also vary. The most widely used standard for gathering and classifying racial and ethnic data by health care systems was adopted from the Office of Management and Budget (OMB) standard created in 1997 for the 2000 U.S. Census. Health Level Seven International, the organization that developed the standard that health care systems use most frequently to send and receive health records, subsequently adopted the OMB classification system from 1997 (see Box 2-3) (Cook et al., 2022). Of note, these standards reflect the 1997 OMB system and may yet be updated per the 2024 OMB revised system.

Many studies have documented that race and ethnicity are frequently missing in EHRs and that, when present, they are often of inconsistent quality (Klinger et al., 2015;

___________________

⁵ https://pcornet.org/ (accessed October 16, 2024).

⁶ https://covid/cd2h.org/about (accessed October 16, 2024).

⁷ https://trinetx.com/ (accessed October 16, 2024).

⁸ https://cosmos.epic.com/ (accessed October 16, 2024).

⁹ https://allofus.nih.gov/ (accessed October 16, 2024).

¹⁰ https://www.mvp.va.gov/ (accessed October 16, 2024).

¹¹ https://fnih.org/observational-medical-outcomes-partnership-omop/ (accessed October 16, 2024).

¹² https://www.enact-network.us/ (accessed October 16, 2024).

¹³ https://covidclinical.net/ (accessed October 16, 2024).

¹⁴ https://ctsi.ucla.edu/uc-health-data-warehouse-uchdw (accessed October 16, 2024).

Polubriaginof et al., 2019). The National COVID Cohort Collaborative examined EHRs from 6.5 million patients from 56 health care institutions in the United States and found that about 21 percent of the race data did not conform to current standards and that about 12 percent of all records were missing race or ethnicity data (Cook et al., 2022). The extent and nature of nonconformance differed according to race and ethnicity, with vulnerable populations and patients of color being disproportionately represented in both the nonconforming and misclassified data. Of note, no patients with race listed as American Indian or Alaska Native (AIAN) were available in the dataset because NIH was in consultation with Tribal leaders and scholars to ensure human protections for research involving the AIAN community and withheld the data of AIAN patients (Cook et al., 2022).

Sources of Biases in EHR Data

Biases in EHR data can stem from a variety of sources; the main sources of biases in EHR data are described below (Chen et al., 2024; Gianfrancesco et al., 2018). Although not unique to EHR data, these potential sources of bias are common in EHR datasets used for biomedical research. Moreover, it is important to recognize bias-driven limitations that could affect study results across racial and ethnic groups. Because EHR data are used routinely in the development, training, and validation of AI algorithms and machine learning (ML) models, failing to correct for the biases in EHR datasets can introduce problems into AI and ML tools.

Selection bias (also known as sampling bias or population bias) arises when the EHR data of individuals or groups used in an analysis are not representative of the larger population, yielding results that cannot be generalized to the larger population. For example, an AI model for forecasting sepsis mortality that is trained on EHR data from a single hospital in a specific geographic region may not generalize well to a broader population such as the entire United States.

Information bias (also known as measurement bias) arises when there are inaccuracies or incompleteness in data entries in the EHR. Incorrect or biased measurements can affect the performance and validity of the analyses. For example, race data in EHRs are acquired through a combination of patient self-report, administrative data entry by health care practitioners, and, in some cases, demographic data transfers from other systems such as insurance databases. Discrepancies in race data in EHRs occur as a result of the heterogeneity of data collection across the 5 Ws (who, what, where, when, and why) (Yemane et al., 2024).

Confounding bias (also known as association bias) arises when a variable, not accounted for in the study design, influences both the predictor and outcome variables, leading to a spurious association between them. In a study predicting patient readmissions, for example, there could be a confounding bias in the data due to socioeconomic status, as those with lower socioeconomic status may have less access to health care resources, contributing to worse medical conditions. In this case, socioeconomic status influences both the input medical problems and the model predictions (Chen et al., 2024). These associations may not be accounted for in EHR data.

Implicit bias refers to unintentional prejudices captured within EHR data due to subjective factors influencing data collection and recording processes. This type of bias may manifest due to health care provider perceptions, patient–provider interactions, or systemic health care practices that influence the way information is recorded in (see, e.g., Himmelstein et al., 2022; Sun et al., 2022). For example, studies have shown that pain assessment and management can vary significantly based on patient demographics such as race (Hoffman et al., 2016). AI models trained on these data can potentially perpetuate disparities in the treatment of pain.

Modeling bias (also known as algorithmic bias) arises when the assumptions, selection of variables, or the design of a model create or amplify the bias in the EHR data. Such bias can occur due to imbalanced or misrepresentative training data, erroneous assumptions made by the model, lack of regulation in model processing, and so on (Chen et al., 2024; Norori et al., 2021). For example, an algorithm that predicts future health care needs based on prior health care costs underestimated the needs of Black patients compared with White patients (Obermeyer et al., 2019).

RACE AND ETHNICITY IN CLINICAL DECISION-MAKING TOOLS

Clinical practice guidelines (CPGs), clinical algorithms, and clinical care pathways are crucial evidence-based practice facilitators. The Institute of Medicine defined CPGs as “statements that include recommendations, intended to optimize patient care, that are informed by a systematic review of evidence and an assessment of the benefits and harms of alternative care options” (IOM, 2011). CPGs guide recommendations for addressing a clinical condition, as they contain a thorough analysis of research evidence, including a benefits and risk assessment for each recommendation; CPGs also include a detailed justification for the recommendation. Often, expert consensus primarily drives recommendations when evidence is insufficient or absent.

Clinical algorithms are typically mathematical formulas, prediction models, flowcharts, or regression equations that assess multiple input variables to discern an outcome probability, such as disease, or a risk estimate of a clinical outcome (Tipton et al., 2023). Investigators employ such algorithms for a variety of clinical purposes (e.g., screening, risk prediction, diagnosis, prognosis, treatment planning, and resource allocation; Table 3-4) (Tipton et al., 2023). Traditional statistical methods (e.g., regression analysis) inform most of the algorithms that clinical investigators currently employ in clinical practice. However, investigators increasingly derive novel algorithms via AI methods, including machine learning.

Clinical care pathways are decision tools that clinicians use to guide evidence-based health care (Busse et al., 2019). Pathways translate clinical practice recommendations into clinical care processes while accounting for the institution’s unique culture and environment. Pathways are often institution-tailored implementations of CPGs or clinical algorithms. However, care teams occasionally predicate their guidelines on data derived exclusively in-house.

TABLE 3-4 Examples of Clinical Algorithms

SOURCE: Content adapted from Visweswaran et al. (2023) CC BY-NC-ND.

Race and Ethnicity in Clinical Practice Guidelines and Clinical Practice Pathways

Many medical professional associations as well as federal agencies, such as the U.S. Centers for Disease Control and Prevention (CDC) and FDA, and volunteer committee-based organizations, such as the U.S. Preventive Services Task Force and the National Academies of Sciences, Engineering, and Medicine, have developed CPGs. Currently, there is scant evidence clarifying the extent to which CPGs incorporate race and ethnicity and any subsequent impact. A systematic review of U.S.-based pediatric CPGs found that race was frequently used in ways that could negatively affect health disparities (Gilliam et al., 2022). The study examined 414 pediatric CPGs and found that 126 (30 percent) incorporated the use of race or ethnicity phrases with 175 occurrences throughout background, recommendations, or future directions (Gilliam et al., 2022). Race was used in a potentially detrimental manner in about 50 percent of instances across 73 CPGs, and in a beneficial way in about 29 percent of instances across 45 CPGs. Potential harmful uses included reinforcing negative stereotypes, conflating race as a biological or genetic risk factor, and normalizing the majority group (Gilliam et al., 2022). Uses of race with potentially positive effects included describing health disparities, inclusivity, and cultural humility.

Even less is known about the extent to which clinical care pathways incorporate race and ethnicity and their resulting effects. A review of clinical care pathways at Boston Children’s Hospital found that 8 (6 percent) out of 132 pathways included race, ethnicity, or ancestry terms (Rosen et al., 2023). Applying a structured framework to evaluate the use of race, ethnicity, and ancestry in these pathways led to the removal or alteration of terms in each instance—in 6 pathways race, ethnicity, or ancestry were eliminated, and 2 pathways were amended (Rosen et al., 2023).

Race and Ethnicity in Clinical Algorithms

When developing clinical algorithms, race and ethnicity are often included based on group-level differences observed in population studies. However, extrapolating

group-level differences based on race or ethnicity to determine an individual’s risk is misguided. For a clinical variable to serve as a predictor of an individual’s risk level, it must be regularly assessed, ordered, and explicitly related to the outcome of interest. For example, the level of LDL (low-density lipoprotein) cholesterol is a variable that is consistently measured in blood tests, it is ordered (increasing concentrations correlate with increasing risk), and it is related to the risk of developing cardiovascular disease. In contrast, race and ethnicity, as well as other social variables, are not consistently measured and are not ordered (e.g., these are nominal variables, with no inherent ranking); the relationship of these variables to clinical outcomes is indirect and ambiguous with many potentially relevant factors.

The rationale to include race and ethnicity as input variables in statistical algorithms is often motivated by observed differences in clinical outcomes between racial and ethnic groups in the studies that developed these algorithms. Vyas et al. (2019), for example, reviewed race-based algorithms in eight clinical specialties and noted this same justification for the inclusion of race and ethnicity as input variables across these different algorithms. However, little, if any, evaluation was done in these studies to assess potential downstream race- or ethnicity-based harms of using such algorithms.

Another persistent challenge with the use of clinical algorithms is that the race and ethnicity values used as inputs for these tools are often heterogenous and unstandardized. For example, one study identified a total of 49 distinct race or ethnicity values in clinical algorithms incorporating race or ethnicity information, the most common of which were White, Black, Other, Asian, Caucasian, East Asian, Mixed, and South Asian (Visweswaran et al., 2023). Almost all of these algorithms use a single race variable, and only a few of them use variables for both race and ethnicity. Some of the algorithms that were examined rely on overly broad race categorizations, such as White/non-White or Black/non-Black, which are inadequate for capturing the full complexities of how patients choose to identify themselves in a clinical context. Some algorithms use just two racial categories, Black or White, and such algorithms are unusable for other research or clinical cases. Thus, these algorithms often employ different sets of race and ethnicity values at varying granularities and with very little consistency, making it challenging to evaluate their current use and develop guidance to inform their future use.

Race and Ethnicity in Clinical AI Algorithms

Clinical algorithms¹⁵ developed using standard statistical methods typically rely on statistical models that are designed to identify patterns and relationships in the data. These models are often based on well-established statistical principles and assumptions, and they typically require researchers to specify the relationship between the variables of interest. In contrast, clinical algorithms developed with AI use computer algorithms that learn relations from the data automatically without being guided by human researchers.

___________________

¹⁵ In the AI literature, algorithms are the methods or procedures used to process data, while models are the outputs of these algorithms that represent learned patterns and relationships. However, in the medical literature, models are frequently referred to as algorithms.

AI is particularly useful for analyzing large, complex datasets with many variables as it can detect complex patterns and relationships in the data that may not be apparent to human researchers. The explosion of data obtained from EHRs as well as data from personal health devices, coupled with advances in computation, is driving the integration of AI throughout the health care ecosystem. AI models are used to predict risk prior to surgery, to assist in emergency department triage, to read medical images, and to evaluate treatment options, along with numerous other uses (Ahmad et al., 2021; Mueller et al., 2022; Rajpurkar et al., 2022). Despite evidence that these models can improve care, there is also growing concern about the presence of bias within these algorithms and their ability to exacerbate existing inequities in the health care system.

While bias in general, and racial bias in particular, is not specific to AI-derived models, the automated nature of AI makes it challenging to evaluate and mitigate bias without explicitly looking for such effects. Bias can enter at any point in the developmental life cycle of an AI model. Moreover, since AI models are often derived from structured and unstructured EHR information that inadequately represents all the demographic factors and social determinants that can affect an individual’s health, the racial and ethnic bias present within these large datasets can be baked in and reinforced by these tools in clinical practice.

The potential for these algorithms to encode and perpetuate bias is exemplified in medical imaging. Current AI technologies, particularly deep learning, are well suited to imaging data, and are increasingly used to process and interpret medical images, such as X-rays, CT scans, and MRIs. AI models can assist in diagnosing conditions, from lung diseases to brain tumors, and have now achieved expert-level performance. However, the models can also display differential performance across subgroups. For example, AI models developed to diagnose pathologies from chest X-rays demonstrated significant underdiagnosis in patients who were Black, female, or of low socioeconomic status (Seyyed-Kalantari et al., 2021).

In addition to radiology, deep learning models have been developed in other medical specialties that use medical images. AI models in dermatology have shown promise in diagnosing skin conditions. However, a systematic review of deep learning models for various skin diseases, including acne, psoriasis, eczema, and rosacea, highlighted the risk of model bias and need for diverse training data (Choy et al., 2023). Others have reported similar limitations in models that are trained predominantly on images of lighter skin tones and less accurate diagnoses for individuals with darker skin (Daneshjou et al., 2022; Groh et al., 2024; Venkatesh et al., 2024), adding to a body of work showing a lack of diverse images of skin tone in dermatology (Alvarado and Feng, 2021; Lester et al., 2020). Similar results have been shown in AI models in ophthalmology (Burlina et al., 2021) and cardiology (Puyol-Antón et al., 2022). Underdiagnosis can result in delayed care for a patient, placing them at higher risk for extended illness, worse outcomes, and higher health care costs.

Multiple studies have found that AI models can develop surprising abilities, such as extracting demographic information (e.g., age, sex, and self-identified race) from medical images (Eng et al., 2021; Gichoya et al., 2022; Yi et al., 2021), even though they are not typically developed to explicitly identify the race or ethnicity of a patient

from a medical image (Gichoya et al., 2022). For instance, Gichoya and colleagues (2022) showed that AI models could accurately detect a patient’s self-reported race from a variety of medical imaging modalities (see also Coyner et al., 2023). How these models can do this remains elusive; no specific features were identified to be responsible for the results. While the ability of an AI model to detect self-reported race from medical images is not meaningful on its own, it is important in the larger context in which AI models are being employed in health care. The apparent ease with which the models learn to identify self-reported race raises the concern that a model may use that information to make race-based interpretations or predictions that could be biased or erroneous, resulting in unfair treatment of patients based on race (Gichoya et al., 2022). This possibility is especially concerning because clinical radiologists examining the same images would be unable to discern race, and without additional information, they would be unable to audit the veracity of the model’s output.

There are several ways that bias could enter the model and contribute to these results. Increasingly, chest X-ray datasets are labeled automatically using natural language processing (NLP) methods (Seyyed-Kalantari et al., 2021). NLP-based methods have been shown to be biased in other health care settings (H. Zhang et al., 2020), and while these labelers have been validated overall, their performance in various subpopulations is untested. Second, the model’s labels are extracted from EHR data, which likely contain biases (see section “Sources of Biases in EHR Data”) that are carried over into the model. This is a form of bias amplification, in which the model outputs amplify biases or errors that exist in the training data (Seyyed-Kalantari et al., 2021). Third, datasets that are too small or lack diversity in racial and ethnic populations may develop biases that favor the dominant populations within the data, leading to inaccurate or unfair predictions for underrepresented groups (Gianfrancesco et al., 2018). Fourth, AI models that extract demographic information such as race and ethnicity may use this information as proxies rather than relying on more relevant factors, resulting in bias (Yang et al., 2024).

Research efforts focused on preventing and removing bias in AI models are emerging throughout the AI and ML community. A variety of fairness metrics have been created to assess fairness across racial and ethnic groups to determine whether an AI model disproportionately penalizes particular racial or ethnic groups (Caton and Haas, 2024). Furthermore, achieving fairness often involves tradeoffs with overall model performance (accuracy) (Caton et al., 2024). If the model is only to be used in a single setting, the best way to obtain a model that is fair while maintaining performance is to optimize for fairness locally (Ghassemi, 2024). Researchers using training data from a single hospital manipulated their model to be more fair and were able to achieve better measures of fairness across demographic groups without a significant loss in model performance (Ghassemi, 2024; Yang et al., 2024). However, these optimization results did not hold when the model was evaluated in a new setting, such as analogous data from a different hospital in a different location. In this scenario where a model will be used in a variety of different settings, prior work has shown that if the only goal is to maximize overall performance, one should select the model with maximum performance on the training domain. However, researchers have shown that this strategy often results in models with degraded fairness during implementation (Ghassemi, 2024; Yang et al., 2024). Instead, researchers have found that selecting models that encode minimal demographic informa-

tion is a more promising strategy that allows for optimal fairness and performance transfer to new hospitals (Ghassemi, 2024; Yang et al., 2024). These case studies demonstrate that reducing race- and ethnicity-based bias associated with AI and ML models and improving patient health outcomes for underserved populations require that different strategies be evaluated and employed based on several factors, meaning that there is no single solution that can address these complex problems. The use of AI in biomedicine and health care is a dynamic, active area of research. Although further exploration was out of scope of this committee,¹⁶ future research will be needed to understand best practices that promote fairness and mitigate performance differences across racial and ethnic groups.

HEALTH DISPARITIES AND THE STUDY OF RACISM

The previous sections describe a number of ways that race and ethnicity are used in biomedical research and thereby affect clinical practice. Given the limitations and biases described, some might wonder what role race and ethnicity serve in biomedical research. This is a question that the committee contemplated over the course of its work. One area the committee examined in its analysis is health disparities. Race and ethnicity have long been used in the study of health disparities and of racism as a driver of persistent health inequity (Heckler, 1985; IOM, 2003; NASEM, 2023, 2024). Racism in health care is evidenced by policies and CPGs that use race and ethnicity as a factor to determine a different standard of care that disadvantages members of racial and ethnic minority groups compared with White individuals. For example, a study on prescribing practices for hypertension medication found that providers were using race-based guidelines to determine the treatment regimen for Black patients, which limited Black patients from receiving the full range of appropriate medications for hypertension (Holt et al., 2022). The study further showed that there was more variation in hypertension control within each racial group (Black and non-Black) than between racial groups (Black versus non-Black) (Holt et al., 2022). Racism in U.S. health policy has also been associated with health disparities experienced by American Indian and Alaska Native people (Solomon et al., 2022). Governmental policies that have “sanctioned inequitable systems of housing, education, employment, health care, environment, and infrastructure” have been associated with lower life expectancy and higher rates of alcohol-related deaths for American Indian and Alaska Native people (Solomon et al., 2022). (See Chapter 5 for more information about structural racism.)

Evidence increasingly indicates that racism, not race, drives health disparities, including, for example, in neuropsychiatry (Carter et al., 2022), asthma (Martinez et al., 2021), and COVID-19 outcomes (Khazanchi et al., 2020; Sabatello et al., 2021). This appears to be in part because racism drives other factors, like socioeconomic status, that influence health. However, socioeconomic status alone does not account for enduring health disparities (Phelan and Link, 2015; Williams et al., 2019b). Research has shown that, independent of socioeconomic status, racism influences health, likely due to inequalities in power, prestige, freedom, neighborhood context, and health care

___________________

¹⁶ For discussion of existing guidance, see also Chapter 4, section “Guidance for Race and Ethnicity in Clinical AI Algorithms.”

(Phelan and Link, 2015). Therefore, racism (defined in Box 2-2) can have direct and indirect effects on health. In addition, racism limits equal access to not only health care but also to participation in biomedical research when beneficial. For instance, Jones (2002 and 2018) notes how racism determines how opportunity is structured, unfairly disadvantaging some individuals and communities while unfairly advantaging others. At the same time, racism appears to have harmful effects on health across racial and ethnic groups, including White populations (Williams et al., 2019a), reducing the health, strength, and potential of the whole society. This evidence contributed to the formulation of the committee’s approach described in Chapter 5 and the development of its recommendations provided in Chapter 6.

CHAPTER SUMMARY AND CONCLUSIONS

Biomedical research comprises many disciplines that include a range of scientific approaches, from basic science to clinical trials to population-level investigations. Thus, the committee examined the common features of the research process as a framework for their analysis. As presented in this chapter, race and ethnicity are commonly used (and sometimes misused) throughout the research process—from study design, to recruitment, analysis, and interpretation of results. The next chapter will build on this framework and assess existing guidance for appropriate use of race and ethnicity in research. This chapter also examined the history of race correction and some of the consequences of this practice that are still seen in science and medicine today. Addressing issues of biased data, misguided approaches, and erroneous assumptions will take effort. Evidence shows that rooting out harmful effects is not simple and that it is important to be wary of unintended consequences. Therefore, based on the evidence presented in this chapter, the committee concluded:

Conclusion 3-1: The incorporation of race and ethnicity into clinical decision-making and care requires nuanced appraisal and consideration to mitigate potential harm to racial and ethnic minority groups and individuals. Issues of race correction or adjustment are not straightforward because removing race can have beneficial, neutral, or harmful effects, which vary within and between racial and ethnic populations and contexts.

Conclusion 3-2: Future efforts to investigate and improve the use of race and ethnicity in clinical algorithms will benefit from a nuanced and context-dependent approach that prioritizes the differential impact that these tools can have on racial and ethnic minority groups who have been harmed by clinical algorithms and other clinical practices.

The evidence presented in this chapter underscored the need for caution in the use of race and ethnicity in biomedical research. The chapter closed with a brief introduction to the evidence from health disparities research that racism, not race, drives health disparities. The persistence of health disparities is a key reason for the continued use of race and ethnicity in research. The following chapter reviews existing guidance on race and ethnicity in biomedical research.

REFERENCES

Brown, O., T. Mou, M. Tate, E. Miller, and M. Debbink. 2022. Considerations for the use of race in research in obstetrics and gynecology. Clinical Obstetrics & Gynecology 65(2):236–243.

Burlina, P., N. Joshi, W. Paul, K. D. Pacheco, and N. M. Bressler. 2021. Addressing artificial intelligence bias in retinal diagnostics. Translational Vision Science & Technology 10(2):13.

Busse, R., N. Klazinga, D. Panteli, and W. Quentin. 2019. Improving healthcare quality in Europe: Characteristics, effectiveness and implementation of different strategies. Health Policy Series, no. 53. Copenhagen: European Observatory on Health Systems and Policies.

Carter, S. E., Y. Mekawi, and N. G. Harnett. 2022. It’s about racism, not race: A call to purge oppressive practices from neuropsychiatry and scientific discovery. Neuropsychopharmacology 47(13):2179–2180.

Carvalho, E. B., T. R. S. Leite, R. F. M. Sacramento, P. Nascimento, C. D. S. Samary, P. R. M. Rocco, and P. L. Silva. 2022. Rationale and limitations of the SpO₂/FiO₂ as a possible substitute for PaO₂/FiO₂ in different preclinical and clinical scenarios. Revista Brasileira de Terapia Intensiva 34(1):185–196.

Caton, S., and C. Haas. 2024. Fairness in machine learning: A survey. ACM Computer Survey 56(7):Article 166.

Cerdeña, J. P., M. V. Plaisime, and J. Tsai. 2020. From race-based to race-conscious medicine: How anti-racist uprisings call us to act. The Lancet 396(10257):1125–1128.

Chen, F., L. Wang, J. Hong, J. Jiang, and L. Zhou. 2024. Unmasking bias in artificial intelligence: A systematic review of bias detection and mitigation strategies in electronic health record-based models. Journal of the American Medical Informatics Association 31(5):1172–1183.

Choy, S. P., B. J. Kim, A. Paolino, W. R. Tan, S. M. L. Lim, J. Seo, S. P. Tan, L. Francis, T. Tsakok, M. Simpson, J. N. W. N. Barker, M. D. Lynch, M. S. Corbett, C. H. Smith, and S. K. Mahil. 2023. Systematic review of deep learning image analyses for the diagnosis and monitoring of skin disease. npj Digital Medicine 6(1):180.

Clinical Algorithms with Race. 2023. Risk calculators, laboratory tests, and therapy recommendations with race-based guidelines. https://clinical-algorithms-with-race.org/risk-calculators-laboratory-tests-therapy-recommendations (accessed March 1, 2024).

CMSS (Council of Medical Specialty Societies). 2023. Reconsidering race in clinical algorithms: Driving equity through new models in research and implementation. Washington, DC.

Cook, L., J. Espinoza, N. G. Weiskopf, N. Mathews, D. A. Dorr, K. L. Gonzales, A. Wilcox, and C. Madlock-Brown. 2022. Issues with variability in electronic health record data about race and ethnicity: Descriptive analysis of the national COVID cohort collaborative data enclave. JMIR Medical Informatics 10(9):e39235.

Coyner, A. S., P. Singh, J. M. Brown, S. Ostmo, R. V. P. Chan, M. F. Chiang, J. Kalpathy-Cramer, J. P. Campbell, B. K. Young, S. J. Kim, K. Sonmez, R. Schelonka, K. Jonas, B. Kolli, J. Horowitz, O. Coki, C.-A. Eccles, L. Sarna, A. Orlin, A. Berrocal, C. Negron, M. K. Denser, K. Cumming, T. Osentoski, T. Check, M. Zajechowski, T. Lee, A. Nagiel, E. Kruger, K. McGovern, D. Contractor, M. Havunjian, C. Simmons, R. Murthy, S. Galvis, J. Rotter, P. I. Chen, X. Li, K. Taylor, K. Roll, M. E. Hartnett, L. Owen, L. Lucci, D. Moshfeghi, M. Nunez, Z. Wennber-Smith, D. Erdogmus, S. Ioannidis, M. A. Martinez-Castellanos, S. Salinas-Longoria, R. Romero, A. Arriola, F. Olguin-Manriquez, M. Meraz-Gutierrez, C. M. Dulanto-Reinoso, and C. Montero-Mendoza. 2023. Association of biomarker-based artificial intelligence with risk of racial bias in retinal images. JAMA Ophthalmology 141(6):543.

Cron, J., A. A. Shapiro, L. Carasimu, J. Iyasere, J. M. Schisler, S. Nagy, S. Angus, A. Burgansky, A. K. Dayal, T. B. Hemmerdinger, D. Howard, C. Oxford-Horrey, D. C. Phillibert, J.-J. Sheen, and D. Goffman. 2024. Understanding clinician knowledge about race adjustment in the vaginal birth after cesarean calculator. Health Equity 8(1):3–7.

Daneshjou, R., K. Vodrahalli, R. A. Novoa, M. Jenkins, W. Liang, V. Rotemberg, J. Ko, S. M. Swetter, E. E. Bailey, O. Gevaert, P. Mukherjee, M. Phung, K. Yekrang, B. Fong, R. Sahasrabudhe, J. A. C. Allerup, U. Okata-Karigane, J. Zou, and A. S. Chiou. 2022. Disparities in dermatology AI performance on a diverse, curated clinical image set. Scientific Advances 8(32):eabq6147.

Delgado, C., M. Baweja, D. C. Crews, N. D. Eneanya, C. A. Gadegbeku, L. A. Inker, M. L. Mendu, W. G. Miller, M. M. Moxey-Mims, G. V. Roberts, W. L. St. Peter, C. Warfield, and N. R. Powe. 2022. A unifying approach for GFR estimation: Recommendations of the NKF–ASN Task Force on Reassessing the Inclusion of Race in Diagnosing Kidney Disease. American Journal of Kidney Diseases 79(2):268–288.e261.

Diao, J. A., Y. He, R. Khazanchi, M. J. Nguemeni Tiako, J. I. Witonsky, E. Pierson, P. Rajpurkar, J. R. Elhawary, L. Melas-Kyriazi, A. Yen, A. R. Martin, S. Levy, C. J. Patel, M. Farhat, L. N. Borrell, M. H. Cho, E. K. Silverman, E. G. Burchard, and A. K. Manrai. 2024. Implications of race adjustment in lung-function equations. New England Journal of Medicine 390(22):2083–2097.

Doris Duke Foundation. 2023. Racial equity in clinical equations. https://www.dorisduke.org/EquityInEquations/ (accessed March 1, 2024).

Drawz, P. E., and J. R. Sedor. 2011. The genetics of common kidney disease: A pathway toward clinical relevance. Nature Reviews Nephrology 7(8):458–468.

Eng, D. K., N. B. Khandwala, J. Long, N. R. Fefferman, S. V. Lala, N. A. Strubel, S. S. Milla, R. W. Filice, S. E. Sharp, A. J. Towbin, M. L. Francavilla, S. L. Kaplan, K. Ecklund, S. P. Prabhu, B. J. Dillon, B. M. Everist, C. G. Anton, M. E. Bittman, R. Dennis, D. B. Larson, J. M. Seekins, C. T. Silva, A. R. Zandieh, C. P. Langlotz, M. P. Lungren, and S. S. Halabi. 2021. Artificial intelligence algorithm improves radiologist performance in skeletal age assessment: A prospective multicenter randomized controlled trial. Radiology 301(3):692–699.

Fatumo, S., T. Chikowore, A. Choudhury, M. Ayub, A. R. Martin, and K. Kuchenbaecker. 2022. Diversity in genomic studies: A roadmap to address the imbalance. Nature Medicine 28(2):243–250.

Feiner, J. R., J. W. Severinghaus, and P. E. Bickler. 2007. Dark skin decreases the accuracy of pulse oximeters at low oxygen saturation: The effects of oximeter probe type and gender. Anesthesia & Analgesia 105(6 Suppl):S18–S23.

FitzGerald, J. D., N. Dalbeth, T. Mikuls, R. Brignardello-Petersen, G. Guyatt, A. M. Abeles, A. C. Gelber, L. R. Harrold, D. Khanna, C. King, G. Levy, C. Libbey, D. Mount, M. H. Pillinger, A. Rosenthal, J. A. Singh, J. E. Sims, B. J. Smith, N. S. Wenger, S. S. Bae, A. Danve, P. P. Khanna, S. C. Kim, A. Lenert, S. Poon, A. Qasim, S. T. Sehra, T. S. K. Sharma, M. Toprover, M. Turgunbaev, L. Zeng, M. A. Zhang, A. S. Turner, and T. Neogi. 2020. 2020 American College of Rheumatology guideline for the management of gout. Arthritis Rheumatology 72(6):879–895.

Freedman, B. I., S. Limou, L. Ma, and J. B. Kopp. 2018. APOL1-associated nephropathy: A key contributor to racial disparities in CKD. American Journal of Kidney Diseases 72(5):S8–S16.

Friedman, D. J., and M. R. Pollak. 2020. APOL1 and kidney disease: From genetics to biology. Annual Review of Physiology 82(1):323–342.

Genzen, J. R., R. J. Souers, L. N. Pearson, D. M. Manthei, A. B. Chambliss, Z. Shajani-Yi, and W. G. Miller. 2023. An update on reported adoption of 2021 CKD–EPI estimated glomerular filtration rate equations. Clinical Chemistry 69(10):1197–1199.

Ghassemi, M. 2024. The pulse of ethical machine learning in multimodal health data. Presentation at Committee on the Use of Race and Ethnicity in Biomedical Research Meeting #3. January 31, 2024.

Gianfrancesco, M. A., S. Tamang, J. Yazdany, and G. Schmajuk. 2018. Potential biases in machine learning algorithms using electronic health record data. JAMA Internal Medicine 178(11):1544–1547.

Gichoya, J. W., I. Banerjee, A. R. Bhimireddy, J. L. Burns, L. A. Celi, L. C. Chen, R. Correa, N. Dullerud, M. Ghassemi, S. C. Huang, P. C. Kuo, M. P. Lungren, L. J. Palmer, B. J. Price, S. Purkayastha, A. T. Pyrros, L. Oakden-Rayner, C. Okechukwu, L. Seyyed-Kalantari, H. Trivedi, R. Wang, Z. Zaiman, and H. Zhang. 2022. AI recognition of patient race in medical imaging: A modelling study. The Lancet Digital Health 4(6):e406–e414.

Gilliam, C. A., E. G. Lindo, S. Cannon, L. Kennedy, T. E. Jewell, and J. S. Tieder. 2022. Use of race in pediatric clinical practice guidelines: A systematic review. JAMA Pediatrics 176(8):804–810.

Goodman, C. W., and A. S. Brett. 2021. Race and pharmacogenomics—Personalized medicine or misguided practice? JAMA 325(7):625–626.

Gottlieb, E. R., J. Ziegler, K. Morley, B. Rush, and L. A. Celi. 2022. Assessment of racial and ethnic differences in oxygen supplementation among patients in the intensive care unit. JAMA Internal Medicine 182(8):849–858.

Grobman, W. A., G. Sandoval, M. M. Rice, J. L. Bailit, S. P. Chauhan, M. M. Costantine, C. Gyamfi-Bannerman, T. D. Metz, S. Parry, D. J. Rouse, G. R. Saade, H. N. Simhan, J. M. Thorp, A. T. N. Tita, M. Longo, and M. B. Landon. 2021. Prediction of vaginal birth after cesarean delivery in term gestations: A calculator without race and ethnicity. American Journal of Obstetrics and Gynecology 225(6):664.e661–664.e667.

Groh, M., O. Badri, R. Daneshjou, A. Koochek, C. Harris, L. R. Soenksen, P. M. Doraiswamy, and R. Picard. 2024. Deep learning-aided decision support for diagnosis of skin disease across skin tones. Nature Medicine 30(2):573–583.

Hammonds, E. M., and R. M. Herzig. 2009. The nature of difference: Sciences of race in the United States from Jefferson to genomics. Cambridge, MA: MIT Press.

Heckler, M. 1985. Report of the Secretary’s Task Force on Black & Minority Health. Washington, DC: U.S. Department of Health and Human Services.

Heffron, A. S., R. Khazanchi, N. Nkinsi, J. A. Bervell, J. P. Cerdeña, J. A. Diao, L. G. Eisenstein, N. J. Gillespie, N. Hongsermeier-Graves, M. Kane, K. Kaur, L. E. Seija, J. Tsai, D. A. Vyas, and A. Y. Zhang. 2022. Trainee perspectives on race, antiracism, and the path toward justice in kidney care. Clinical Journal of the American Society of Nephrology 17(8):1251–1254.

Himmelstein, G., D. Bates, and L. Zhou. 2022. Examination of stigmatizing language in the electronic health record. JAMA Network Open 5(1):e2144967.

Hoenig, M. P., A. Mann, and M. Pavlakis. 2022. Removal of the Black race coefficient from the estimated glomerular filtration equation improves transplant eligibility for Black patients at a single center. Clinical Transplantation 36(2):e14467.

Hoffman, K. M., S. Trawalter, J. R. Axt, and M. N. Oliver. 2016. Racial bias in pain assessment and treatment recommendations, and false beliefs about biological differences between blacks and whites. Proceedings of the National Academy of Sciences 113(16):4296–4301.

Holder, A. L., and A. I. Wong. 2022. The big consequences of small discrepancies: Why racial differences in pulse oximetry errors matter. Critical Care Medicine 50(2):335–337.

Holt, H. K., G. Ginny, K. Leah, F. Valy, P. Rajiv, and B. P. Michael. 2022. Differences in hypertension medication prescribing for Black Americans and their association with hypertension outcomes. Journal of the American Board of Family Medicine 35(1):26–34.

Igo, S. E. 2007. The averaged American: Surveys, citizens, and the making of a mass public. Cambridge, MA: Harvard University Press.

Inker, L. A., S. J. Couture, H. Tighiouart, A. G. Abraham, G. J. Beck, H. I. Feldman, T. Greene, V. Gudnason, A. B. Karger, J. H. Eckfeldt, B. L. Kasiske, M. Mauer, G. Navis, E. D. Poggio, P. Rossing, M. G. Shlipak, and A. S. Levey. 2021. A new panel-estimated GFR, including β(2)-microglobulin and β-trace protein and not including race, developed in a diverse population. American Journal of Kidney Diseases 77(5):673–683.e671.

IOM (Institute of Medicine). 1993. Iron deficiency anemia: Recommended guidelines for the prevention, detection, and management among U.S. children and women of childbearing age. Washington, DC: National Academy Press.

IOM. 2003. Unequal treatment: Confronting racial and ethnic disparities in health care. Washington, DC: The National Academies Press.

IOM. 2009. Initial national priorities for comparative effectiveness research. Washington, DC: The National Academies Press.

IOM. 2011. Clinical practice guidelines we can trust. Washington, DC: The National Academies Press.

Jakachira, R., M. Diouf, Z. Lin, J. A. Burrow, A. Howes, T. Oguntolu, R. Carter III, S. I. Dunsiger, and K. C. Toussaint Jr. 2022. Single-wavelength, single-shot pulse oximetry using an led-generated vector beam. Optics Express 30(15):27293–27303.

Jones, C., G. McQuillan, J. Kusek, M. Eberhardt, W. Herman, J. Coresh, M. Salive, C. Jones, and L. Agodoa. 1998. Serum creatinine levels in the U.S. population: Third National Health and Nutrition Examination Survey. American Journal of Kidney Diseases 32(6): 992–999.

Jones, C. P. 2002. Confronting institutionalized racism. Phylon (1960-) 50(1/2):7.

Jones, C. P. 2018. Toward the science and practice of anti-racism: Launching a national campaign against racism. Ethnicity & Disease 28(Supp 1):231.

Khan, S. S., J. Coresh, M. J. Pencina, C. E. Ndumele, J. Rangaswami, S. L. Chow, L. P. Palaniappan, L. S. Sperling, S. S. Virani, J. E. Ho, I. J. Neeland, K. R. Tuttle, R. Rajgopal Singh, M. S. V. Elkind, and D. M. Lloyd-Jones for the American Heart Association, 2023. Novel prediction equations for absolute risk assessment of total cardiovascular disease incorporating cardiovascular–kidney–metabolic health: A scientific statement from the American Heart Association. Circulation 148(24):1982–2004.

Khazanchi, R., C. T. Evans, and J. R. Marcelin. 2020. Racism, not race, drives inequity across the COVID-19 continuum. JAMA Network Open 3(9):e2019933.

Klinger, E. V., S. V. Carlini, I. Gonzalez, S. S. Hubert, J. A. Linder, N. A. Rigotti, E. Z. Kontos, E. R. Park, L. X. Marinacci, and J. S. Haas. 2015. Accuracy of race, ethnicity, and language preference in an electronic health record. Journal of General Internal Medicine 30(6):719–723.

Landy, R., I. Gomez, T. J. Caverly, K. Kawamoto, M. P. Rivera, H. A. Robbins, C. D. Young, A. K. Chaturvedi, L. C. Cheung, and H. A. Katki. 2023. Methods for using race and ethnicity in prediction models for lung cancer screening eligibility. JAMA Network Open 6(9): e2331155.

Lester, J. C., J. L. Jia, L. Zhang, G. A. Okoye, and E. Linos. 2020. Absence of images of skin of colour in publications of COVID-19 skin manifestations. British Journal of Dermatology 183(3):593–595.

Levey, A. S., J. P. Bosch, J. B. Lewis, T. Greene, N. Rogers, D. Roth, and Modification of Diet in Renal Disease Study Group. 1999. A more accurate method to estimate glomerular filtration rate from serum creatinine: A new prediction equation. Annals of Internal Medicine 130(6):461–470.

Lujan, H. L., and S. E. DiCarlo. 2018. Science reflects history as society influences science: Brief history of “race,” “race correction,” and the spirometer. Advances in Physiology Education 42(2):163–165.

Martinez, A., R. De La Rosa, M. Mujahid, and N. Thakur. 2021. Structural racism and its pathways to asthma and atopic dermatitis. Journal of Allergy and Clinical Immunology 148(5):1112–1120.

McFarling, U. L. 2022. “A poster child” for diversity in science: Black engineers work to fix long-ignored bias in oxygen readings. https://www.statnews.com/2022/08/19/diversity-in-science-black-engineers-work-to-fix-long-ignored-bias-in-pulse-oximeters/ (accessed April 8, 2024).

Mensah, G. A. 2019. Black–white disparities: More than just race. Journal of the American Heart Association 8(22):e014272.

Merid, B., and V. Volpe. 2023. Race correction and algorithmic bias in atrial fibrillation wearable technologies. Health Equity 7(1):817–824.

Merz, L. E., C. M. Story, M. A. Osei, K. Jolley, S. Ren, H. S. Park, R. Yefidoff Freedman, D. Neuberg, R. Smeland-Wagman, R. M. Kaufman, and M. O. Achebe. 2023. Absolute neutrophil count by Duffy status among healthy Black and African American adults. Blood Advances 7(3):317–320.

Moran-Thomas, A. 2020. How a popular medical device encodes racial bias. https://www.bostonreview.net/articles/amy-moran-thomas-pulse-oximeter/ (accessed April 8, 2024).

Mueller, B., T. Kinoshita, A. Peebles, M. A. Graber, and S. Lee. 2022. Artificial intelligence and machine learning in emergency medicine: A narrative review. Acute Medical Surgery 9(1):e740.

NASEM (National Academies of Sciences, Engineering, and Medicine). 2023. Using population descriptors in genetics and genomics research: A new framework for an evolving field. Washington, DC: The National Academies Press.

NASEM. 2024. Ending unequal treatment: Strategies to achieve equitable health care and optimal health for all. Washington, DC: The National Academies Press.

Nead, K. T., C. L. Hinkston, and M. R. Wehner. 2022. Cautions when using race and ethnicity in administrative claims data sets. JAMA Health Forum 3(7):e221812.

Nicholas, S. B., K. Kalantar-Zadeh, and K. C. Norris. 2015. Socioeconomic disparities in chronic kidney disease. Advances in Chronic Kidney Disease 22(1):6–15.

NIDDK (National Institute of Diabetes and Digestive and Kidney Diseases). 2023. Kidney disease statistics for the United States. https://www.niddk.nih.gov/health-information/health-statistics/kidney-disease (accessed July 3, 2024).

NIH (National Institutes of Health). 2017. Does your human subjects research study meet the NIH definition of a clinical trial? https://grants.nih.gov/ct-decision/index.htm (accessed July 1, 2024).

Norori, N., Q. Hu, F. M. Aellen, F. D. Faraci, and A. Tzovara. 2021. Addressing bias in big data and AI for health care: A call for open science. Patterns 2(10):100347.

Norton, J. M., M. M. Moxey-Mims, P. W. Eggers, A. S. Narva, R. A. Star, P. L. Kimmel, and G. P. Rodgers. 2016. Social determinants of racial disparities in CKD. Journal of the American Society of Nephrology 27(9):2576–2595.

O’Brien, T. J., K. Fenton, A. Sidahmed, A. Barbour, and A. F. Harralson. 2021. Race and drug toxicity: A study of three cardiovascular drugs with strong pharmacogenetic recommendations. Journal of Personalized Medicine 11(11):1226.

O’Brien, J., and C. A. Clare. 2023. Race-based versus race-conscious medicine in obstetrics and gynecology. Clinical Obstetrics & Gynecology 66(1):95–106.

Obermeyer, Z., B. Powers, C. Vogeli, and S. Mullainathan. 2019. Dissecting racial bias in an algorithm used to manage the health of populations. Science 366(6464):447–453.

Office of the National Coordinator for Health Information Technology. 2023. National trends in hospital and physician adoption of electronic health records. https://www.healthit.gov/data/quickstats/national-trends-hospital-and-physician-adoption-electronic-health-records (accessed April 8, 2024).

Okunlola, O. E., M. S. Lipnick, P. B. Batchelder, M. Bernstein, J. R. Feiner, and P. E. Bickler. 2022. Pulse oximeter performance, racial inequity, and the work ahead. Respiratory Care 67(2):252–257.

Patwari, N., D. Huang, and F. Bonetta-Misteli. 2024. Short: Racial disparities in pulse oximetry cannot be fixed with race-based correction. Paper presented at Proceedings of the 8th ACM/IEEE International Conference on Connected Health: Applications, Systems and Engineering Technologies, Orlando, FL, USA.

Peterson, P. N., J. S. Rumsfeld, L. Liang, N. M. Albert, A. F. Hernandez, E. D. Peterson, G. C. Fonarow, and F. A. Masoudi. 2010. A validated risk score for in-hospital mortality in patients with heart failure from the American Heart Association get with the guidelines program. Circulation: Cardiovascular Quality and Outcomes 3(1):25–32.

Phelan, J. C., and B. G. Link. 2015. Is racism a fundamental cause of inequalities in health? Annual Review of Sociology 41(1):311–330.

Polubriaginof, F. C. G., P. Ryan, H. Salmasian, A. W. Shapiro, A. Perotte, M. M. Safford, G. Hripcsak, S. Smith, N. P. Tatonetti, and D. K. Vawdrey. 2019. Challenges with quality of race and ethnicity data in observational databases. Journal of the American Medical Informatics Association 26(8–9):730–736.

Pottel, H., J. Björk, A. D. Rule, N. Ebert, B. O. Eriksen, L. Dubourg, E. Vidal-Petiot, A. Grubb, M. Hansson, E. J. Lamb, K. Littmann, C. Mariat, T. Melsom, E. Schaeffner, P. O. Sundin, A. Åkesson, A. Larsson, E. Cavalier, J. B. Bukabau, E. K. Sumaili, E. Yayo, D. Monnet, M. Flamant, U. Nyman, and P. Delanaye. 2023. Cystatin C-based equation to estimate GFR without the inclusion of race and sex. New England Journal of Medicine 388(4): 333–343.

Powe, N. R. 2020. Black kidney function matters: Use or misuse of race? JAMA 324(8): 737–738.

Powe, N. R. 2021. The pathogenesis of race and ethnic disparities: Targets for achieving health equity. Clinical Journal of the American Society of Nephrology 16(5):806–808.

Powe, N. R. 2022. Race and kidney function: The facts and fix amidst the fuss, fuzziness, and fiction. Med 3(2):93–97.

Puyol-Antón, E., B. Ruijsink, J. Mariscal Harana, S. K. Piechnik, S. Neubauer, S. E. Petersen, R. Razavi, P. Chowienczyk, and A. P. King. 2022. Fairness in cardiac magnetic resonance imaging: Assessing sex and racial bias in deep learning-based segmentation. Frontiers in Cardiovascular Medicine 9:859310.

Rajpurkar, P., E. Chen, O. Banerjee, and E. J. Topol. 2022. AI in health and medicine. Nature Medicine 28(1):31–38.

Reese, P. P., N. R. Powe, and B. Lo. 2024. Engineering equity into the promise of xenotransplantation. American Journal of Kidney Diseases 83(5):677–683.

Rosen, R. H., A. Epee-Bounya, D. Curran, S. Chung, R. Hoffmann, L. K. Lee, C. Marcus, C. M. Mateo, J. E. Miller, C. Nereim, E. Silberholz, S. N. Shah, C. V. Theodoris, H. Wardell, A. S. Winn, S. Toomey, J. A. Finkelstein, V. L. Ward, and A. Starmer. 2023. Race, ethnicity, and ancestry in clinical pathways: A framework for evaluation. Pediatrics 152(6):e2022060730.

Sabatello, M., M. Jackson Scroggins, G. Goto, A. Santiago, A. McCormick, K. J. Morris, C. R. Daulton, C. L. Easter, and G. Darien. 2021. Structural racism in the COVID-19 pandemic: Moving forward. American Journal of Bioethics 21(3):56–74.

Seyyed-Kalantari, L., H. Zhang, M. B. A. McDermott, I. Y. Chen, and M. Ghassemi. 2021. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature Medicine 27(12):2176–2182.

Shahian, D. M., J. P. Jacobs, V. Badhwar, P. A. Kurlansky, A. P. Furnary, J. C. Cleveland, Jr., K. W. Lobdell, C. Vassileva, M. C. Wyler von Ballmoos, V. H. Thourani, J. S. Rankin, J. R. Edgerton, R. S. D’Agostino, N. D. Desai, L. Feng, X. He, and S. M. O’Brien. 2018. The Society of Thoracic Surgeons 2018 adult cardiac surgery risk models: Part 1—Background, design considerations, and model development. Annals of Thoracic Surgery 105(5):1411–1418.

Siddique, S. M., K. Tipton, B. Leas, C. Jepson, J. Aysola, J. B. Cohen, E. Flores, M. O. Harhay, H. Schmidt, G. E. Weissman, J. Fricke, J. R. Treadwell, and N. K. Mull. 2024. The impact of health care algorithms on racial and ethnic disparities: A systematic review. Annals of Internal Medicine 177(4):484–496.

Sjoding, M. W., R. P. Dickson, T. J. Iwashyna, S. E. Gay, and T. S. Valley. 2020. Racial bias in pulse oximetry measurement. New England Journal of Medicine 383(25):2477–2478.

Sjoding, M. W., T. J. Iwashyna, and T. S. Valley. 2023. Change the framework for pulse oximeter regulation to ensure clinicians can give patients the oxygen they need. American Journal of Respiratory and Critical Care Medicine 207(6):661–664.

Solomon, T. G. A., R. R. B. Starks, A. Attakai, F. Molina, F. Cordova-Marks, M. Kahn-John, C. L. Antone, M. Flores, and F. Garcia. 2022. The generational impact of racism on health: Voices from American Indian communities. Health Affairs 41(2):281–288.

Stanojevic, S., D. A. Kaminsky, M. R. Miller, B. Thompson, A. Aliverti, I. Barjaktarevic, B. G. Cooper, B. Culver, E. Derom, and G. L. Hall. 2022. ERS/ATS technical standard on interpretive strategies for routine lung function tests. European Respiratory Journal 60(1):2101499.

Stevens, L. A., J. Coresh, C. H. Schmid, H. I. Feldman, M. Froissart, J. Kusek, J. Rossert, F. Van Lente, R. D. Bruce, 3rd, Y. L. Zhang, T. Greene, and A. S. Levey. 2008. Estimating GFR using serum cystatin C alone and in combination with serum creatinine: A pooled analysis of 3,418 individuals with CKD. Am J Kidney Dis 51(3):395–406.

Sun, M., T. Oliwa, M. E. Peek, and E. L. Tung. 2022. Negative patient descriptors: Documenting racial bias in the electronic health record. Health Affairs 41(2):203–211.

Tipton, K., B. F. Leas, E. Flores, C. Jepson, J. Aysola, J. Cohen, M. Harhay, H. Schmidt, G. Weissman, J. Treadwell, N. K. Mull, and S. M. Siddique. 2023. Impact of healthcare algorithms on racial and ethnic disparities in health and healthcare. Comparative Effectiveness Review no. 268. Rockville, MD: Agency for Healthcare Research and Quality.

Turner, B. E., J. R. Steinberg, B. T. Weeks, F. Rodriguez, and M. R. Cullen. 2022. Race/ethnicity reporting and representation in U.S. clinical trials: A cohort study. The Lancet Regional Health–Americas 11:100252.

Valbuena, V. S. M., S. Seelye, M. W. Sjoding, T. S. Valley, R. P. Dickson, S. E. Gay, D. Claar, H. C. Prescott, and T. J. Iwashyna. 2022. Racial bias and reproducibility in pulse oximetry among medical and surgical inpatients in general care in the Veterans Health Administration, 2013–19: Multicenter, retrospective cohort study. BMJ 378:e069775.

Valley, T., A. Moran-Thomas, and F. Tessema. 2023. Discriminating devices: The case of pulse oximetry. https://www.youtube.com/watch?v=8d8cL1pZPyc (accessed April 8, 2024).

Varughese, P. M., L. Krishnan, and Ravichandran. 2018. Does color really matter? Reliability of transcutaneous bilirubinometry in different skin-colored babies. Indian Journal of Paediatric Dermatology 19(4):315–320.

Venkatesh, K. P., M. M. Raza, G. Nickel, S. Wang, and J. C. Kvedar. 2024. Deep learning models across the range of skin disease. npj Digital Medicine 7(1):32.

Vesoulis, Z., A. Tims, H. Lodhi, N. Lalos, and H. Whitehead. 2022. Racial discrepancy in pulse oximeter accuracy in preterm infants. Journal of Perinatology 42(1):79–85.

Visweswaran, S., E. M. Sadhu, M. M. Morris, and M. J. Samayamuthu. 2023. Clinical Algorithms with Race: An online database. medRxiv. Preprint July 6, 2023. https://doi.org/10.1101/2023.07.04.23292231.

Vyas, D. A., D. S. Jones, A. R. Meadows, K. Diouf, N. M. Nour, and J. Schantz-Dunn. 2019. Challenging the use of race in the vaginal birth after cesarean section calculator. Women’s Health Issues 29(3):201–204.

Vyas, D. A., L. G. Eisenstein, and D. S. Jones. 2020. Hidden in plain sight—Reconsidering the use of race correction in clinical algorithms. New England Journal of Medicine 383(9):874–882.

Williams, D. R., J. A. Lawrence, and B. A. Davis. 2019a. Racism and health: Evidence and needed research. Annual Review of Public Health 40105–40125.

Williams, D. R., J. A. Lawrence, B. A. Davis, and C. Vu. 2019b. Understanding how discrimination can affect health. Health Services Research 54(Suppl 2):1374–1388.

Wong, A. I., M. Charpignon, H. Kim, C. Josef, A. A. H. de Hond, J. J. Fojas, A. Tabaie, X. Liu, E. Mireles-Cabodevila, L. Carvalho, R. Kamaleswaran, R. Madushani, L. Adhikari, A. L. Holder, E. W. Steyerberg, T. G. Buchman, M. E. Lough, and L. A. Celi. 2021. Analysis of discrepancies between pulse oximetry and arterial oxygen saturation measurements by race and ethnicity and association with organ dysfunction and mortality. JAMA Network Open 4(11):e2131674.

Wright, J. L., W. S. Davis, M. M. Joseph, A. M. Ellison, N. J. Heard-Garris, T. L. Johnson, and the American Academy of Pediatrics Board Committee on Equity. 2022. Eliminating race-based medicine. Pediatrics 150(1):e2022057998.

Yang, Y., H. Zhang, J. W. Gichoya, D. Katabi, and M. Ghassemi. 2024. The limits of fair medical imaging AI in real-world generalization. Nature Medicine 30(10):2838–2848.

Yearby, R. 2020. Structural racism and health disparities: Reconfiguring the social determinants of health framework to include the root cause. Journal of Law, Medicine & Ethics 48(3):518–526.

Yemane, L., C. M. Mateo, and A. N. Desai. 2024. Race and ethnicity data in electronic health records—Striving for clarity. JAMA Network Open 7(3):e240522.

Yi, P. H., J. Wei, T. K. Kim, J. Shin, H. I. Sair, F. K. Hui, G. D. Hager, and C. T. Lin. 2021. Radiology “forensics”: Determination of age and sex from chest radiographs using deep learning. Emergency Radiology 28(5):949–954.

Zhang, X., T. A. Melanson, L. C. Plantinga, M. Basu, S. O. Pastan, S. Mohan, D. H. Howard, J. M. Hockenberry, M. D. Garber, and R. E. Patzer. 2018. Racial/ethnic disparities in waitlisting for deceased donor kidney transplantation 1 year after implementation of the new national kidney allocation system. American Journal of Transplantation 18(8):1936–1946.

Zou, J., and L. Schiebinger. 2021. Ensuring that biomedical AI benefits diverse populations. EBioMedicine 67:103358.

Zuberi, T. 2003. Thicker than blood: How racial statistics lie. Minneapolis, MN: University of Minnesota Press.

This page intentionally left blank.