Tag: Racial Bias

When “Neutral” Data Isn’t: Gender and Racial Bias in Biomedical AI

Bias in AI isn’t abstract. It shows up in real diagnoses, real treatment decisions, and real patients, particularly those who already face multiple layers of disadvantage. Discrimination in AI is an intersectional issue, overlapping with age, gender, race, sexual orientation, and health status, among other dimensions of identity. The patients most at risk of being harmed by biased AI tools are often those who already face the most barriers to quality care.

Decades of biased medical data

The biases we see in AI didn’t come from nowhere. They’re a direct consequence of medical data collected over decades, data shaped by systemic discrimination and structural inequalities in healthcare research and practice (Cirillo et al., 2020; Cross et al., 2024).

For much of medical history, clinical and experimental studies focused predominantly on male participants. This is what is known as the gender health gap. In 2020, only 5% of global health research funding went to women’s health research. This was split into 4% for women’s cancers and 1% for all other women-specific health conditions, with 25% of that further limited to fertility research (Nature Reviews Bioengineering, 2024).

It’s worth noting this isn’t a one-way problem: in some areas, such as depression, men are underrepresented in clinical data, largely because they’re less likely to seek care, report symptoms, or receive a diagnosis (Smith et al., 2018).

Racial bias is equally well documented

Racial bias in medicine is also extensively documented, particularly in the United States where research shows that, for instance, Black patients and other minority groups receive fewer medical procedures, lower rates of surgical intervention, and fewer referrals to specialists than white patients, regardless of clinical need (Bowser, 2001; Williams & Wyatt, 2015). When AI systems are trained on data that reflects decades of existing inequalities in healthcare access and treatment, they risk encoding and perpetuating those inequalities at a much larger scale. While much of the evidence comes from the US, the structural conditions producing racial bias, including socioeconomic inequalities, barriers to access and underrepresentation in research, are present across Europe too.

LGBTQIA+ patients face their own layer of risk

It is also important to recognise the specific situation of LGBTQIA+ individuals, who experience discrimination in healthcare and are subject to stereotypes that affect the care they receive. These social and cultural factors perpetuate discrimination and have a measurable impact on health and healthcare. Research has shown that 16% of LGBTQIA+ individuals report discrimination in healthcare encounters, and 18% avoid seeking care altogether due to fear of mistreatment (Chang et al., 2025). Another large-scale study found that in emergency department scenarios, AI recommendations for LGBTQIA+ patients included mental health interventions six to seven times more often than was clinically appropriate (Chang et al., 2025).

How this plays out in three key areas

In cardiovascular care, women’s symptoms often differ from the “classic” presentation described in medical textbooks, itself largely based on male patient data (Fatunde et al., 2025). As a result, women are offered fewer diagnostic tests, less medication, and fewer specialist referrals (Al Hamid et al., 2024). Racial bias compounds the problem: pulse oximeters, devices routinely used to measure blood oxygen saturation, have been shown to produce less accurate readings for patients with darker skin tones (Sjoding et al., 2020), a bias that then feeds directly into AI-based triage systems.

In diabetes care, racial bias is extensively documented. Studies have shown that African American patients present systematically higher A1c values than white patients with the same average blood glucose (Karter et al., 2023). If an AI system uses A1c as a proxy for glycaemic control without accounting for this, it risks producing incorrect diagnoses for Black patients.

In depression, both gender and racial bias carry significant weight, especially in tools using natural language processing for screening. Men and women tend to express psychological distress differently (Pennebaker et al., 2003), so a system trained predominantly on one gender’s language patterns may screen inaccurately for the other. Most mental health AI tools also still operate on binary gender assumptions, excluding non-binary, transgender, and gender non-conforming individuals from both the data and the populations these tools are meant to serve (Hafner et al., 2024).

Why this matters

These biases translate into delayed diagnoses, inappropriate treatments, and unequal care, every day, for real patients. Understanding these patterns is exactly what equips healthcare professionals to ask better questions, challenge assumptions, and advocate for the patients most at risk, which is precisely what the AEQUITAS training is designed to do.

References:

Chang, C.T., Srivathsa N., Bou-Khalil, C., Swaminathan, A., Lunn, M.R., Mishra, K., Koyejo, S,. Daneshjou, R. (2025). Evaluating anti-LGBTQIA+ medical bias in large language models. PLOS Digit Health 4(9): e0001001. https://doi.org/10.1371/journal.pdig.0001001

Cirillo, D., Catuara-Solarz, S., Morey, C., Guney, E., Subirats, L., Mellino, S., Gigante, A.A., Valencia, A., Rementeria, M.J., Chadha, A.S., & Mavridis, N. (2020). Sex and gender differences and biases in artificial intelligence for biomedicine and healthcare. npj Digit. Med. 3(81). https://doi.org/10.1038/s41746-020-0288-5

Cross, J. L., Choma, M. A., & Onofrey, J. A. (2024). Bias in medical AI: Implications for clinical decision-making. PLOS Digital Health, 3(11), e0000651. https://doi.org/10.1371/journal.pdig.0000651

Funding research on women’s health. (2024). Nature Reviews Bioengineering, 2, 797–798. https://doi.org/10.1038/s44222-024-00253-7

Hafner, F.S., Valdivia, A., Rocher, L. 2025. Gender Trouble in Language Models: An Empirical Audit Guided by Gender Performativity Theory. FAccT ’25: Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, 1677–1695. https://doi.org/10.1145/3715275.3732112

Karter, A. J., Parker, M. M., Moffet, H. H., & Gilliam, L. K. (2023). Racial and Ethnic Differences in the Association Between Mean Glucose and Hemoglobin A1c. Diabetes Technology & Therapeutics, 25(10), 697–704. https://doi.org/10.1089/dia.2023.0153

Pennebaker, J. W., Mehl, M. R., & Niederhoffer, K. G. (2003). Psychological Aspects of Natural Language Use: Our Words, Our Selves. Annual Review of Psychology, 54(1), 547–577. https://doi.org/10.1146/annurev.psych.54.101601.145041

Sjoding, M. W., Dickson, R. P., Iwashyna, T. J., Gay, S. E., & Valley, T. S. (2020). Racial Bias in Pulse Oximetry Measurement. New England Journal of Medicine, 383(25), 2477–2478. https://doi.org/10.1056/NEJMc2029240

Smith, D.T., Mouzon, D.M., & Elliott, M. (2018). Reviewing the assumptions about men’s mental health: An exploration of the gender binary. American Journal of Men’s Health, 12(1), 78–89. https://doi.org/10.1177/1557988316630953

Bowser, R. (2001). Racial bias in medical treatment. Dick. L. Rev., 105(3), 365.

Williams, D. R., & Wyatt, R. (2015). Racial Bias in Health Care and Health: Challenges and Opportunities. JAMA, 314(6), 555. https://doi.org/10.1001/jama.2015.9260
The bias algorithm: how AI in healthcare exacerbates ethnic and racial disparities – a scoping review

AI health algorithms are increasingly used to predict risk and support treatment decisions, often outperforming humans in diagnostic accuracy. However, they can also underestimate risk for marginalised and underserved patients, delaying the treatment they need. This scoping review analysed 23 sources published between 2020 and 2024 to examine racial and ethnic bias in these algorithms, the role of oversight and stakeholders, and how this bias can be reduced.

A common explanation for this bias is that algorithms are trained on data in which majority populations are overrepresented, making them less reliable for others. Black, Native American, Asian and Latinx populations are often underrepresented in health datasets. Bias can also come from people: racial stereotypes absorbed during medical training can affect how clinicians assess patients, and these assessments end up in the health records used to train algorithms. The authors also describe a feedback loop: when patients from marginalised groups distrust the health system and use it less, data about them becomes even less representative, and the algorithms less accurate for them.

To address this, the review highlights the need for continuous monitoring of algorithms once they are in use, in a similar way to how medicines are monitored after approval, and for retraining or withdrawing those that create inequities. It also stresses that simply including or removing race from an algorithm does not solve the problem, and that responsibility should be shared by developers, health systems, clinicians, regulators and patients. The authors conclude that, without this oversight, AI in healthcare risks entrenching health inequities.

Learn more about this review here: https://doi.org/10.1080/13557858.2024.2422848

Reference:

Hussain, S.A., Bresnahan, M., & Zhuang, J. (2025). The bias algorithm: how AI in healthcare exacerbates ethnic and racial disparities – a scoping review. Ethnicity & Health, 30(2), 197–214.
Addressing hidden risks: Systematic review of artificial intelligence biases across racial and ethnic groups in cardiovascular diseases

AI models are increasingly used in cardiovascular medicine, but there are concerns that they may not work equally well for people of different racial and ethnic backgrounds. This systematic review examined how accurate and clinically applicable AI models for diagnosing heart disease and predicting outcomes are across racial and ethnic groups.

Of 1,704 references screened, 11 studies were included, covering diagnosis, prognosis and the analysis of heart images. Nine of them (82%) found racial or ethnic bias, while two found no differences between groups. For example, an AI model predicting heart failure from electrocardiograms was less accurate for young Black patients than for young Asian or White patients, and AI tools analysing cardiac MRI scans were more accurate for White people than for people from minority ethnic groups.

The authors conclude that AI models can perform well for some groups and less accurately for others, and that bias should be addressed at every stage, from data collection to clinical use. They recommend methods to reduce bias, standardised ways of measuring fairness, considering social determinants of health rather than race alone, and greater transparency through the sharing of code and data.

Learn more about this review here: https://doi.org/10.1016/j.ejrad.2024.111867

Reference:

Cau, R., Pisu, F., Suri, J.S., & Saba, L. (2024). Addressing hidden risks: Systematic review of artificial intelligence biases across racial and ethnic groups in cardiovascular diseases. European Journal of Radiology, 111867.
Artificial Intelligence to Promote Racial and Ethnic Cardiovascular Health Equity

AI has been seen as a tool that could help detect and reduce the racial and ethnic inequities that persist in cardiovascular care. However, AI can also carry and spread biases of its own, which means it must be approached with care. This review looks at the role of AI in cardiology in identifying and addressing these inequities, and in promoting diversity in research, as well as the concerns its use raises and strategies for using it fairly.

The authors describe how AI has already helped reduce some inequities. In one study, an AI tool identified patients with severe heart valve disease and alerted their doctors, improving follow-up care for patients of all races. Even so, Black patients still faced greater delays in receiving appropriate care than White patients, showing that AI is not a “silver bullet” for eliminating inequities. AI can also perpetuate existing biases, due to limited diversity in training data, biases within the datasets themselves, and inadequate mechanisms for detecting and monitoring bias. For example, an AI model trained mostly on data from White patients to predict heart failure from electrocardiograms performed worse for Black patients, and particularly for Black women.

To use AI in cardiovascular care in a way that addresses racial and ethnic inequities, the authors stress the need for careful design, starting with meticulous data collection and a thorough review of training data. They also call for rigorous testing and continuous monitoring of algorithms to ensure they perform equally well for all groups, as well as greater diversity in the AI workforce and the involvement of stakeholders.

Learn more about this review here: https://doi.org/10.1007/s12170-024-00745-6

Reference:

Amponsah, D., Thamman, R., Brandt, E., James, C., Spector-Bagdady, K., & Yong, C.M. (2024). Artificial Intelligence to Promote Racial and Ethnic Cardiovascular Health Equity. Current Cardiovascular Risk Reports, 18(11), 153–162.
Algorithmic Bias in AI-Based Diabetes Care: Systematic Review of Model Performance, Equity Reporting, and Physiological Label Bias

AI has transformed diabetes management, particularly through tools that predict glucose levels and support clinical decisions. However, there are concerns that these technologies may not work equally well for different racial, ethnic and socioeconomic groups. This systematic review assessed how fair AI-based glucose prediction models are, focusing on differences in performance between groups, how fairness is reported, and biases linked to physiological measures.

Of 1,243 articles initially identified, only 10 met the criteria for inclusion. Just one of them explicitly evaluated whether AI model performance differed by race. Reporting on fairness was rare: only 7% of AI diabetes studies documented ethnic or racial data, and almost none carried out fairness audits. The review also highlights known physiological differences: for the same average glucose level, African American patients tend to have higher HbA1c values than White patients. If HbA1c is used as a reference to train AI models, this difference can be built into them. Yet none of the AI studies reviewed took this into account.

The authors also note that traditional diabetes risk scores overestimate risk for White individuals and underestimate it for Black individuals, a bias that AI systems trained on these scores could inherit. They conclude that, without systematic efforts to reduce bias, AI-based diabetes technologies risk worsening existing health disparities, and call for future research to prioritise transparency, representativeness and fairness.

Learn more about this review here: https://doi.org/10.61186/ist.202502.05.04

Reference:

Alipour, M., & Alipour, A. (2025). Algorithmic Bias in AI-Based Diabetes Care: Systematic Review of Model Performance, Equity Reporting, and Physiological Label Bias. InfoScience Trends, 2(5), 33–46.
Racial disparities in continuous glucose monitoring-based 60-min glucose predictions among people with type 1 diabetes

Continuous glucose monitors, small sensors worn on the body that track blood sugar levels throughout the day, are increasingly combined with AI models that predict where glucose levels are heading. Because White people are overrepresented in medical research, these models may perform worse for other groups if they are trained mainly on White patients’ data. This study tested this using data from 205 adults and children with type 1 diabetes in the US, 101 White and 104 Black.

The researchers trained AI models to predict glucose levels 60 minutes ahead, using datasets with different proportions of White and Black participants, from 0% to 100% White. As the proportion of White participants in the training data increased, prediction errors grew slightly more for Black participants than for White participants. The difference was small but statistically significant.

When the models were then personalised to each individual, using a technique called transfer learning, this difference became smaller and was no longer statistically significant. The authors note that the differences may not be large enough to matter clinically on their own, but conclude that diverse training data is important for fair prediction models, and that personalisation may help build fairer tools.

Learn more about this study here: https://doi.org/10.1371/journal.pdig.0000918

Reference:

Thomsen, H.B., Li, L.Y., Isaksen, A.A., Lebiecka-Johansen, B., Bour, C., Fagherazzi, G., van Doorn, W.P.T.M., Varga, T.V., & Hulman, A. (2025). Racial disparities in continuous glucose monitoring-based 60-min glucose predictions among people with type 1 diabetes. PLOS Digital Health, 4(6), e0000918.
Real-world artificial intelligence-based opportunistic screening for diabetic retinopathy in endocrinology and indigenous healthcare settings in Australia

Diabetic retinopathy, damage to the retina caused by diabetes, can lead to vision loss if it is not detected early. This study tested an AI system for diabetic retinopathy screening in real-world conditions, in two endocrinology clinics in Melbourne and three Aboriginal medical services in Western Australia, and also gathered the views of patients and health professionals.

Of the 236 people screened, 203 were included in the main analysis. Overall, the system identified diabetic retinopathy requiring referral with a sensitivity of 96.9% and a specificity of 87.7%, and over 93% of participants were satisfied with the service. However, specificity was substantially lower among participants recruited from Aboriginal medical services, meaning that more people without the condition were wrongly flagged for referral. The study used clinic location as an approximation of Indigenous status.

The authors note that the AI system had been trained exclusively on images from Chinese adults, and that differences in retinal pigmentation may affect its accuracy. They also point to other possible factors, such as poorer image quality, older age and other eye conditions. These explanations were not tested in the study.

Learn more about this study here: https://doi.org/10.1038/s41598-021-94178-5

Reference:

Scheetz, J., Koca, D., McGuinness, M., Holloway, E., Tan, Z., Zhu, Z., O’Day, R., Sandhu, S., MacIsaac, R.J., Gilfillan, C., Turner, A., Keel, S., & He, M. (2021). Real-world artificial intelligence-based opportunistic screening for diabetic retinopathy in endocrinology and indigenous healthcare settings in Australia. Scientific Reports, 11, 15808.
Assessing and Mitigating Bias in Medical Artificial Intelligence: The Effects of Race and Ethnicity on a Deep Learning Model for ECG Analysis

AI models developed using data from very similar populations may not work as well for other groups, and could reflect or even worsen existing racial and ethnic disparities in health care. This study tested whether this was the case for an AI model that uses electrocardiograms (ECGs) to detect a weakened heart pumping function, known as low left ventricular ejection fraction.

The model had been developed with data from nearly 45,000 patients, 96.2% of whom were non-Hispanic White. When tested on a separate group of almost 53,000 patients, it performed consistently well across all the racial and ethnic groups analysed: non-Hispanic White, Asian, Black/African American, Hispanic/Latino and American Indian/Native Alaskan. A separate AI model was able to identify patients’ racial group from the ECG to some extent, showing that ECG characteristics do vary by race, but this did not affect the first model’s performance.

The authors conclude that, in this case, racial differences in the ECG did not affect the model’s ability to detect the condition. Even so, they recommend that the performance of all new AI tools in medicine be reported across different ethnic, racial, age and sex groups, to ensure they are used responsibly.

Learn more about this study here: https://doi.org/10.1161/CIRCEP.119.007988

Reference:

Noseworthy, P.A., Attia, Z.I., Brewer, L.C., Hayes, S.N., Yao, X., Kapa, S., Friedman, P.A., & Lopez-Jimenez, F. (2020). Assessing and Mitigating Bias in Medical Artificial Intelligence: The Effects of Race and Ethnicity on a Deep Learning Model for ECG Analysis. Circulation: Arrhythmia and Electrophysiology, 13(3), e007988.