Artificial Intelligence as a Tool to Reduce Racial Bias in Medicine: Challenges and Opportunities

Health care workers, such as physicians, physician associates, and nurse practitioners, are susceptible to implicit racial biases that can influence clinical decision making. These biases may present as shorter consultation times, delays in or omission of diagnostic testing, and suboptimal disease management among people of color (POC).

A recent study found that provider visits with Black patients were, on average, nearly 2.5 minutes shorter than those with White patients.1 Although this difference may appear minimal, it is troubling in high-volume clinical environments, where every minute of patient interaction can affect the quality of care.

Physicians are also statistically less likely to order diagnostic imaging, such as computed tomography (CT), MRI, and ultrasonography, for non-White patients.2,3 These missed opportunities for timely diagnosis can lead to repeated emergency department visits, delayed diagnosis and treatment, and poorer health outcomes. Although often subtle and unintentional, such disparities perpetuate the health care inequities experienced by POC.

Racism in Medical Training Programs

Although providers’ implicit bias contributes to inequitable patient care, medical training programs and medical literature are part of the larger system that sustains these disparities. From their inception, Western medical schools were shaped by scientific racism, which influenced institutional standards and curricula. At the first medical school in the United States, later known as the University of Pennsylvania,4 the influential physician Samuel George Morton published research that advanced White supremacist ideologies. These views were incorporated into medical education at the time, influencing generations of physicians and contributing to biased treatment of patients.

During that period, much of the medical research promoted the false narrative that White individuals were inherently superior, citing purported differences such as larger brain size or categorizing White and Black individuals as genetically distinct species. Although these ideologies have evolved, their effects continue to shape modern medicine.

One enduring stereotype is the false belief that Black people have a higher tolerance for pain.5 This misconception often leads to the under prescription of pain medications and the damaging assumption that Black patients are more likely to be “drug-seeking.”5 The legacy of scientific racism continues to permeate modern health care, shaping biased clinical practices and influencing medical decision-making standards.

Commonly Missed Diagnoses in POC

Race-based decision making persists in health care despite the fact that race does not represent the complex pathophysiology of many disease processes.6 Educational materials used to train health care providers often center on White patients, with limited representation of clinical findings relevant to POC. In dermatology, for example, diagnostic criteria for skin lesions are predominantly based on presentations in lighter skin tones. As a result, studies consistently demonstrate that Black patients experience delayed diagnosis of melanoma, contributing to lower survival rates compared with their White counterparts.7

A widely accepted laboratory test in nephrology, the estimated glomerular filtration rate (eGFR), historically incorporated a race-based adjustment. For more than 2 decades, the calculation included a modifier for Black patients, which often delayed diagnosis of chronic kidney disease in this population and contributed to disparities in timely care and disease management.8

Racial disparities are also prevalent in obstetrics and gynecology, where pregnant Black women face disproportionately higher risks for adverse outcomes. In fact, Black women face a 4-fold higher risk for death from pregnancy-related complications compared with White women, a disparity influenced in part by the persistent false belief that they have a higher pain tolerance.5

These examples, although not exhaustive, illustrate how the legacy of scientific racism continues to shape inequities in the diagnosis and treatment of POC within the US health care system.

Artificial intelligence (AI) has become a more widespread fixture in the health care system. As health care and technology continue to advance, their integration has become increasingly intertwined. The application of AI in health care is not new; it initially emerged through efforts to automate medical diagnostics.9 Over time, AI has advanced considerably and is now used across a wide range of functions, such as predictive modeling, clinical decision support, and administrative tasks.

AI algorithms can analyze complex datasets such as genetic profiles, radiologic images, and electronic health records, with a level of speed and accuracy that surpasses human capability.10 AI has demonstrated superior performance compared with radiologists in both speed and accuracy for detection of tumors on CT scans and identification of fractures on radiographs.10 In addition, AI is increasingly integrated into early warning systems, such as systems that flag abnormal electrocardiograms (ECGs) to facilitate faster clinical response.11

AI integration can also enhance service flexibility within hospitals. Service flexibility encompasses both volume flexibility and clinical flexibility, each of which directly affects patient satisfaction, care delivery, and outcomes.12 Volume flexibility refers to a hospital’s capacity to adjust resources and services in response to patient volume. AI supports this function by forecasting patient flow, anticipating disease outbreaks, and identifying shifts in health care demand, thereby allowing more streamlined and effective care.12

AI has the potential to transform modern health care; however, without intentional and inclusive design, it risks perpetuating the racial disparities already entrenched in the US health care system.

Clinical flexibility refers to the ability to diagnose patients quickly and accurately. AI models can rapidly analyze a wide range of patient data, such as medical history, diagnostic imaging, and test results, enabling earlier and more precise diagnosis. Early disease detection supports the development of more comprehensive treatment plans, which may lead to improved patient outcomes. By leveraging AI to enhance clinical adaptability, hospitals can reduce costs, improve operational efficiency, and maintain a strong focus on high-quality patient care.12

Training AI tools requires large datasets from which algorithms identify patterns. This process allows programs to “learn” and draw conclusions or make decisions on the basis of the available data.13 When datasets lack diversity or fail to represent all populations accurately, AI tools can inherit and reproduce these biases, potentially resulting in discriminatory outcomes.14 Because access to health care remains disproportionately limited for POC, these populations often experience delayed diagnosis and higher mortality rates. Consequently, the progression and presentation of many diseases are less comprehensively documented in these populations than in White individuals. This lack of diverse data can lead AI systems to overlook early signs of disease on diagnostic imaging, perpetuating existing gaps in care.

Machine learning, for example, is being used in dermatology to develop AI tools that can distinguish benign from malignant skin lesions.15 The dataset used by the International Skin Imaging Collaboration remains disproportionately representative of individuals with lighter skin tones. Relying on a tool of this magnitude without addressing such bias risks reinforcing existing disparities in the detection of malignant lesions in darker skin. Until datasets become more inclusive and reflective of the broader population, AI will continue to perform optimally primarily for the individuals represented in the data.

Black box decision making refers to AI algorithms whose internal decision-making processes are not transparent to developers or users.16 AI can process information at a level that exceeds human cognitive capacity. The processes through which AI interprets data and generates conclusions are often opaque and difficult to explain. In health care, when a model is trained on skewed or nonrepresentative data, its performance often declines among patient populations that are not adequately represented in the dataset.17 The lack of transparency in AI decision-making makes potential inaccuracies difficult for providers to recognize, and without careful oversight, these errors can adversely affect patient outcomes.

Furthermore, although AI is often regarded as objective and data-driven, it remains susceptible to biases embedded within its training data. This may result in clinicians depending on AI output without recognizing underlying bias, which is more difficult to identify when AI generates unexplainable decisions.18 These decisions are made with limited visibility, restricting the ability to detect errors and hindering efforts to refine and improve these models.19

Discussion: How to Improve AI to Reduce Bias

AI continues to be one of the most impressive technological advances in recent years, particularly in health care. Significant limitations need to be addressed for its true potential to be realized. One of the most effective ways to improve AI is to increase the diversity of the data used to train these models. Because of long-standing disparities in access to medical treatment, gathering this information remains a major challenge. A coordinated nationwide effort across institutions would be needed to generate datasets diverse enough to train models that can accurately serve the population as a whole. Without inclusive datasets, AI systems risk perpetuating bias and reinforcing the health inequities that they could instead help address.19

In 2022, the National Institutes of Health (NIH) launched the Bridge2AI program to generate AI-ready data aimed at addressing gaps in care and improving treatment for underrepresented populations.14 Although evidence on the program’s effect remains preliminary, other studies have shown that retraining AI models can enhance predictive performance and potentially improve health outcomes for marginalized groups.20

Promoting interdisciplinary collaboration is imperative for building equitable and effective AI models. Recent research indicates that the development of fair and effective models in health care requires input from diverse disciplines, such as clinicians, ethicists, and data scientists.21 Although software developers and data scientists possess the technical expertise to build models that accurately reflect the data provided, a model created solely from that perspective may lack the clinical insight required to ensure that its outputs are medically valid and relevant. Clinicians help guide these models toward practical application, whereas ethicists can strengthen them further by addressing unintended consequences, such as the misrepresentation of vulnerable populations. Incorporating diverse perspectives throughout the development process allows potential blind spots to be identified and addressed early, promoting both clinical accuracy and social relevance.

The controversy surrounding bias in AI models is not exclusively a technological issue. The current health care system is constructed in ways that contribute to and may exacerbate this problem. Health care systems need policies, standards, and continuous oversight to ensure equitable use of AI.22 Through programs such as Bridge2AI, the NIH requires that AI models be trained on demographically representative and ethically sourced data. This equity-driven effort can establish a standard for diverse and ethical data practices. Institutions such as the NIH and FDA can also require a minimal level of model interpretability, helping reduce the black box decision making that obscures underlying biases. Implementing these measures helps maintain core medical ethical principles like beneficence and justice.23 Regular audits to evaluate AI model performance across diverse demographics, such as race, gender, and age, can also help identify and prevent the deployment of biased systems.24 These continuous assessments can further monitor the efficacy of AI models throughout each stage of development.25