Tag: Healthcare

Sociodemographic biases in medical decision making by large language models

Large language models (LLMs), such as the AI behind popular chatbots, show promise in healthcare, but they may make recommendations influenced by patients’ sociodemographic characteristics. Researchers tested nine LLMs on 1,000 emergency department cases, half real and half synthetic. Each case was presented in 32 versions: one for each of 31 sociodemographic groups, plus a control version with no sociodemographic details. The clinical details were always the same. In total, they analysed more than 1.7 million AI-generated recommendations.

Compared with both a physician-defined baseline and the control cases, cases labelled as Black or unhoused or as LGBTQIA+ were more often directed towards urgent care, invasive interventions or mental health evaluations. For example, certain cases from LGBTQIA+ subgroups were recommended mental health assessments around six to seven times more often than clinically indicated. Cases labelled as high-income were more often recommended advanced imaging tests, such as CT or MRI, while low- and middle-income cases were often limited to basic tests or none at all.

These differences were not supported by clinical reasoning or guidelines and may reflect bias in the models, which could lead to health disparities. As they were found in both proprietary and open-source models, the authors stress the need for robust ways to evaluate and reduce bias, so that AI-driven medical advice remains fair and centred on patients.

Learn more about this study here: https://doi.org/10.1038/s41591-025-03626-6

Reference:

Omar, M., Soffer, S., Agbareia, R., Bragazzi, N. L., Apakama, D. U., Horowitz, C. R., Charney, A. W., Freeman, R., Kummer, B., Glicksberg, B. S., Nadkarni, G. N., & Klang, E. (2025). Sociodemographic biases in medical decision making by large language models. Nature Medicine, 31, 1873–1881.
Gender biases within Artificial Intelligence and ChatGPT: Evidence, Sources of Biases and Solutions

AI is now used in many areas of our lives, from hiring and education to healthcare, but it often mirrors the prejudices that exist in society. This review brings together evidence on how AI systems, and chatbots such as ChatGPT in particular, can perpetuate gender bias. Examples include a hiring algorithm that learned to favour men, and ChatGPT writing recommendation letters that described women with more social words, such as “family”, while letters for men emphasised their professional skills.

In healthcare, the review explains that many AI tools are trained on data that comes mostly from men, which can lead to unequal care. Women’s heart disease symptoms are more likely than men’s to be misdiagnosed or attributed to other conditions, even when they report the same symptoms, and AI trained on this kind of data can repeat these patterns. The review also cites studies showing that AI tools for diagnosing chest diseases underdiagnosed women more often, and that GPT-4 was more likely to recommend treatments and further tests for men than for women.

The authors identify three main sources of bias: the data used to train AI, the way the systems are designed, and the feedback loop created by users’ interactions with them. To address these, they propose improving the quality and diversity of data and of the people who label it, using algorithms designed with fairness in mind, and strengthening policies at company, national and international levels, such as the EU AI Act.

Learn more about this review here: https://doi.org/10.1016/j.chbah.2025.100145

Reference:

Ho, J. Q. H., Hartanto, A., Koh, A., & Majeed, N. M. (2025). Gender biases within artificial intelligence and ChatGPT: Evidence, sources of biases and solutions. Computers in Human Behavior: Artificial Humans, 4, 100145.
FDA proposes updated recommendations to help improve performance of pulse oximeters across skin tones

Pulse oximeters are small devices, usually placed on a finger, that estimate the level of oxygen in the blood. Evidence shows they may be less accurate in people with darker skin. This problem gained wider attention during the COVID-19 pandemic, when inaccurate readings led to delayed treatment for some patients with darker skin.

In January 2025, the US Food and Drug Administration (FDA) proposed draft guidance with updated recommendations for pulse oximeters used for medical purposes, mainly in hospitals and doctors’ offices. Manufacturers should collect clinical data to assess how accurate their devices are across different skin tones. They should also include more study participants, so that studies better reflect the patients who will use the device, and use standardised methods to classify participants’ skin tone, such as the Monk Skin Tone scale.

The FDA also recommends that labels clearly state any limitations of the device for different skin tones, and it plans a public website listing cleared pulse oximeters. The guidance does not cover devices sold for general wellness or sports. Together, these measures aim to help patients and health professionals identify devices that work reliably for everyone.

Learn more here: https://www.fda.gov/news-events/press-announcements/fda-proposes-updated-recommendations-help-improve-performance-pulse-oximeters-across-skin-tones

Reference:

U.S. Food and Drug Administration. (2025, January 6). FDA proposes updated recommendations to help improve performance of pulse oximeters across skin tones [Press announcement].
Turning Evidence Into Action: The AEQUITAS Database

AI tools used in healthcare can carry hidden biases, sometimes performing less accurately for women, ethnic minorities, or other underrepresented groups. But knowing this risk exists in general is one thing. Being able to check whether a specific tool, in a specific clinical area, has documented bias issues is another, and that’s exactly the gap the AEQUITAS Database was built to fill.

The AEQUITAS Database is a structured, digital repository that systematically collects, organises, and provides access to evidence on gender and racial bias in biomedical AI. Rather than leaving healthcare professionals, researchers, and policymakers to piece together scattered studies and reports, it brings this dispersed evidence together into one coherent, accessible knowledge base.

The database is publicly accessible through the AEQUITAS project website, and it was built with a user-oriented approach, meaning both technical and non-technical users can navigate it effectively.

Built for multiple audiences

The Database was designed to serve several groups at once, each with different needs. For healthcare professionals, it’s a resource for identifying potential biases in AI-based diagnostic and decision-support tools, improving patient-centred care, and reducing the risk of misdiagnosis. For civil society organisations, it supports awareness-raising and advocacy for fair, inclusive healthcare systems. And for researchers and policymakers, it feeds into further scientific investigation, the development of clinical guidelines, and regulatory oversight at European level.

A practical tool

The Database is designed to be used, not just consulted occasionally. Before adopting a new AI tool, healthcare professionals can search for entries related to its disease area to check whether bias has been documented in similar tools. When an AI system produces unexpected outputs, the Database can help determine whether it’s a known pattern rather than an isolated anomaly. And when caring for patients from underrepresented groups, it can help clinicians understand whether the tools involved in their care carry documented disparities relevant to that patient’s profile.

The Database is also a living resource that will grow as new evidence becomes available. The more healthcare professionals, researchers, and organisations engage with it, the more valuable it becomes for everyone.
Effect of Ethnicity on HbA1c Levels in Individuals without Diabetes: Systematic Review and Meta-Analysis

HbA1c is a blood test that reflects average blood sugar levels over the previous months, and it is widely used to diagnose diabetes using a single cut-off value for everyone. However, HbA1c levels appear to differ between ethnic groups. Because most earlier studies included people with diabetes, whose blood sugar varies more, this systematic review focused only on people without diabetes.

The review combined the results of 12 studies with a total of 49,238 adults. Compared with White participants, HbA1c levels were significantly higher in Black participants (by 0.26 percentage points), Asian participants (by 0.24 points) and Latino participants (by 0.08 points).

The authors conclude that, although these differences are small, they may affect the use of a single HbA1c threshold to diagnose diabetes in all ethnic groups.

Learn more about this review here: https://doi.org/10.1371/journal.pone.0171315

Reference:

Cavagnolli, G., Pimentel, A.L., Freitas, P.A.C., Gross, J.L., & Camargo, J.L. (2017). Effect of Ethnicity on HbA1c Levels in Individuals without Diabetes: Systematic Review and Meta-Analysis. PLOS ONE, 12(2), e0171315.
Race, Racism, and Risk Prediction for Cardiovascular Disease

To estimate a patient’s risk of heart disease and stroke, many clinicians have used a tool called the Pooled Cohort Equations, which had separate equations for Black and White people. In recent years, concerns have grown that this approach may reinforce harmful stereotypes and suggest that race itself has a biological basis for disease. In this commentary, two cardiologists from the American Heart Association explain why the association’s new risk calculator, PREVENT, no longer uses race.

Instead of race, PREVENT considers factors such as kidney and metabolic health, which affect Black Americans more often, and a social deprivation index based on postcode, reflecting the social conditions that drive racial disparities in cardiovascular disease. It was developed and validated using data from more than 3 million people each, with diverse races and ethnicities and equal numbers of women and men, and proved accurate across all racial and ethnic groups studied. According to the authors, leaving race out also helps to make clear that race is associated with disease, but does not cause it.

The authors stress, however, that a more accurate calculator does not eliminate unconscious bias or the injustices that people of certain races and ethnicities face in healthcare. It is the effects of racism, not race itself, that are linked to disease risk. They describe PREVENT as a first step towards fairer tools in cardiovascular care, and note that research is under way to test whether it works equally well for everyone, regardless of how they identify, how much they earn or where they live.

Learn more here: https://newsroom.heart.org/news/race-racism-and-risk-prediction-for-cardiovascular-disease

Reference:

Churchwell, K., & Khan, S.S. (2024, March 12). Race, Racism, and Risk Prediction for Cardiovascular Disease. American Heart Association Newsroom.
Addressing bias: artificial intelligence in cardiovascular medicine

AI offers opportunities to transform cardiovascular medicine, but AI algorithms can also have biases of their own. In this commentary, the authors describe several sources of bias in cardiovascular AI. These include differences in how diagnostic tests such as echocardiograms and cardiac MRI are carried out and interpreted, models that learn patterns from their training data that do not hold in new populations, and the biases that clinicians may bring, consciously or not, when labelling the data used to train them.

A particular concern is that AI could worsen health disparities. Women and minority groups have historically been under-represented in cardiology research, so much of the current evidence may not apply to them. The authors give examples of inequalities that end up in the data used to train AI: men are 2.5 times more likely than women to be referred to a cardiologist for chest pain, and Black patients in emergency departments are 40% less likely to receive pain medication than White patients. In another example, an algorithm identified patients’ postcodes, linked to low-income and predominantly Black neighbourhoods, as the main predictor of a longer hospital stay.

To prevent this, the authors call for clinicians, ethicists and representatives of minority groups to be involved from the early stages of AI development, for algorithms to be validated in minority and low-income groups, and for regulatory and ethical guidelines to ensure that the benefits of AI are shared equally, along with regular monitoring of AI models once they are in use.

Learn more about this article here: https://doi.org/10.1016/S2589-7500(20)30249-1

Reference:

Tat, E., Bhatt, D.L., & Rabbat, M.G. (2020). Addressing bias: artificial intelligence in cardiovascular medicine. The Lancet Digital Health, 2(12), e635–e636.
From Principles to Practice: How AEQUITAS Turns Policy into Action

AI systems used in healthcare are governed by strict ethical principles and legal requirements, from the EU AI Act to the EU Charter of Fundamental Rights. But knowing the rules isn’t the same as knowing how to apply them. Hospitals and civil society organisations (CSOs) need concrete, operational processes they can actually put into practice, day to day, to make sure these tools are used safely and fairly.

This is exactly what the AEQUITAS AI Regulatory Model was designed to do. It gives hospitals and CSOs a practical governance framework covering the full lifecycle of AI systems, from procurement and validation to deployment, monitoring, and accountability.

The model draws inspiration from approaches that have proven effective in other high-risk sectors. It echoes the logic of pharmacovigilance, the systematic tracking of adverse drug reactions, in what AEQUITAS calls an “AI-vigilance” system: a way to report and analyse incidents involving AI-supported clinical decisions, with a particular focus on equity outcomes. It also mirrors the patient safety culture familiar from aviation-inspired approaches in hospitals, built around identifying hazards, reporting incidents, and continuously improving.

At its core, the model is grounded in the EU Charter of Fundamental Rights (2009), particularly the principles of human dignity, equality and non-discrimination, and the right to healthcare, translating these rights into operational safeguards that follow an AI system throughout its use.

The five checkpoints

Rather than treating AI governance as a single box to tick, the model structures oversight around five sequential stages, each one a checkpoint designed to catch potential equity risks before they reach patients. Throughout this process, hospitals act as the main deployers, leading decisions and operational control, while CSOs play an independent oversight role, reviewing documentation, flagging equity-related risks, and conducting audits, without controlling day-to-day clinical decisions themselves.

These stages follow the natural lifecycle of any AI system in a hospital. It starts with scope and classification, determining an AI tool’s regulatory status before it’s even procured or developed. This sets the direction for everything that follows. Next comes procurement, often the moment hospitals have the greatest leverage, as it’s where suppliers can be required to commit to addressing bias risks throughout a system’s lifecycle. At this stage, CSOs are able to review procurement documentation and flag equity concerns. The third stage, validation, means independently evaluating how a system performs, not just overall, but across different patient groups, before it’s ever used with real patients.

Once a system passes validation, the deployment stage ensures the right safeguards are in place: trained oversight staff, mechanisms for clinicians to override AI recommendations, and clear communication with patients about when these tools are being used in their care. And because AI systems aren’t static, the final stage, ongoing monitoring, auditing, and accountability, keeps watching for emerging disparities long after a system goes live, with CSOs conducting independent audits and helping escalate persistent issues.

Building a shared network across Europe

Beyond the model itself, AEQUITAS is currently establishing a European network of CSOs and public hospitals. This shared governance layer means that insights, incidents, and mitigation strategies identified in one country contribute to a growing, Europe-wide knowledge base, one that benefits patients well beyond the institution where the issue was first spotted.

Together, these tools turn abstract principles into something hospitals and CSOs can actually act on, closing the gap between good intentions and equitable, everyday care.

References:

European Union. (2009). Charter of Fundamental Rights of the European Union. Official Journal of the European Union, C 303/1. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:12012P/TXT
Improving Health in Alzheimer’s Patients
Proposal: Adapted Physical Activity and Awareness Training to Improve Health in Alzheimer’s Patients and Their Entourage

Proposed Implementation: 2024 to 2026

Call: EU4H-2023-PJ – EU4H Action Grants 2023

Proposed Budget: 596 289,60€

Keywords: Alzheimer, Physical Activity Practices, Physical Activity, Health, Inequality

Objective: Dementia is one of the most significant public health concerns in the European Union with around 10 million people living with this condition. The prevalence of Alzheimer’s disease is expected to double in the next 30 years with the increase in the aging population. As a consequence, a part of the population will risk becoming dependent on carers and specific socio-medical services.

As a result, this will increase health costs for the EU. The project’s overall objective is to contribute to prevent dependency, thus reducing the expenditure of the European healthcare system on Alzheimer disease and other neurological cognitive impairments by providing to patients who have been diagnosed with Alzheimer and/or dementia with adapted physical activity practices to help slow down the progress of the disease and manage their symptoms.

To achieve these objectives, actions will focus on people who have been diagnosed with dementia at a pre-symptomatic early stage and with MCI-minor cognitive impairment. First, this will be achieved through a person-centered physical activity programme with the support of functional ability training and guidelines for a healthier lifestyle. It will support 500 patients by means of 50 person-centered physical activity sessions over the course of one year across 3 European countries.

Second, the project aims to raise awareness on physical activity as a tool to manage neurological diseases and as a tool to prevent the progression of cognitive impairment symptoms (such as memory loss, problems communicating etc.) and to create an integrated care model for dementia patients with a 3-staged approach 1st test, 2nd evaluate and 3rd implement. The project will provide a sustainable model by constant evaluation process and sharing the results across the European member states.

Third, an online “Train the Trainer” programme will provide a foundation for further upskilling of health care practitioners, physical activity trainers and caregivers.

Partners:
- Fundacion Siel Bleu España
- Health Citizens – European Institute
- Siel Bleu Ireland
- Sport Initiative et Loisir Bleu Association
- Alzheimer Catalunya Fundació
Beyond Ethics: How Europe Is Regulating AI in Healthcare

Knowing that AI can be biased is one thing. Knowing what to do about it, legally and ethically, is another.

Medicine’s ethical foundations, extended to AI

Medicine has long been associated with high ethical standards, from the Hippocratic Oath to the Declarations of Geneva and Helsinki. Unfortunately, the history of medicine includes deeply troubling chapters, from eugenics, forced sterilisation of indigenous and marginalised populations, to non-consensual experimentation on people with disabilities. These historical failures are directly relevant today because the structural injustices behind them shaped decades of medical data collection, and their legacy is exactly what today’s AI systems risk inheriting when trained on that same data.

Biomedical ethics is built around four core ethical principles: autonomy (respecting a person’s right to make their own decisions), non-maleficence (avoiding harm), beneficence (actively promoting wellbeing), and justice (distributing benefits and risks fairly, especially given known disparities based on race, gender, and social status) (Beauchamp & Childress, 2019).

As AI entered clinical practice, ethicists added a fifth principle specifically for these technologies: explicability, the idea that the reasoning behind an AI system’s decisions must be understandable and open to scrutiny, not simply accepted at face value (Floridi et al., 2018).

The EU’s legal response

The European Union has built a multi-layered legal framework specifically to address these risks. At its centre is the AI Act, which classifies many healthcare applications, including diagnostic tools and clinical decision-support systems, as “high-risk.” This means they’re subject to strict requirements: providers must actively identify and mitigate risks of discriminatory outcomes, ensure meaningful human oversight, and maintain accuracy and transparency throughout the system’s lifecycle (EU AI Act, 2024).

Hospitals that deploy these systems have specific legal obligations too, including using AI in line with provider instructions, assigning competent human oversight, maintaining logs of system operation, and reporting serious incidents, particularly where there’s a risk of unequal outcomes across patient groups.

The AI Act doesn’t stand alone. It works alongside the Medical Device Regulation (MDR; European Union, 2017a) and In Vitro Diagnostic Regulation (IVDR; European Union, 2017b), which govern the safety of medical technologies, the GDPR (European Union, 2016), which protects the sensitive health data these systems rely on, and the emerging European Health Data Space (EHDS; European Union, 2025), designed to enable more representative datasets for future AI development. Together, they form a legal ecosystem, all ultimately grounded in the EU Charter of Fundamental Rights (2009), particularly its guarantees of non-discrimination, equality, and the right to healthcare.

All these policies translate into real responsibilities for hospitals and healthcare professionals, from understanding a tool’s limitations and knowing when to override its recommendations, to being transparent with patients about when and how these systems are used in their care.

Beyond raising awareness, the AEQUITAS project has developed a practical AI Regulatory Model that translates these legal and ethical principles into concrete steps hospitals and healthcare professionals can actually apply, turning complex EU legislation into something usable in everyday clinical practice.

References:

Beauchamp, T. L., & Childress, J. F. (2019). Principles of biomedical ethics (8th ed).

European Commission, Timeline for Implementation of the EU AI Act, Brussels: European Commission, 2024.

European Union. (2009). Charter of Fundamental Rights of the European Union. Official Journal of the European Union, C 303/1. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:12012P/TXT

European Union, Regulation (EU) 2016/679 on General Data Protection Regulation (GDPR), Brussels: EU, 2016.

European Union, Regulation (EU) 2017/745 on Medical Devices (MDR), Official Journal of the European Union, L117, Brussels: European Union, 2017.

European Union, Regulation (EU) 2017/746 on In Vitro Diagnostic Medical Devices (IVDR), Official Journal of the European Union, L117, Brussels: European Union, 2017.

European Union, Regulation (EU) 2025/327 on the European Health Data Space (EHDS), Brussels: European Union, 2025.

Floridi, L., Cowls, J., Beltrametti, M., Chatila, R., Chazerand, P., Dignum, V., Luetge, C., Madelin, R., Pagallo, U., Rossi, F., Schafer, B., Valcke, P., & Vayena, E. (2018). AI4People—An Ethical Framework for a Good AI Society: Opportunities, Risks, Principles, and Recommendations. Minds and Machines, 28(4), 689–707. https://doi.org/10.1007/s11023-018-9482-5
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 Hidden Ways AI Can Discriminate

When we talk about bias in AI, we are referring to “computer systems that systematically and unfairly discriminate against certain individuals or groups in favour of others” (Friedman & Nissenbaum, 1996). This is not about occasional errors, but about consistent, predictable patterns of unfair outcomes. Furthermore, biases in AI systems are complex, as they can enter the system at almost any stage, from the data it’s trained on to the way it’s ultimately used in a hospital setting. Recognising where bias originates, and how different types can reinforce each other, is the first step towards addressing it.

Three broad roots of bias

Researchers have identified three broad categories of bias that apply to all computer systems. The first is pre-existing bias, which comes from existing inequalities in society that get absorbed into the system, sometimes without anyone realising it (Friedman & Nissenbaum, 1996). The second is technical bias, which arises from the practical compromises made when translating complex human realities into clean computational form. And the third is emergent bias, which appears only after a system is deployed, as the world around it changes in ways the original design never anticipated.

Bias in the AI pipeline

These three categories give us a useful starting point. But for AI and machine learning specifically, researchers have identified seven more precise bias types that can affect these systems throughout their development and use (Suresh & Guttag, 2021):

Historical bias reflects prejudices and stereotypes already present in training data, even if the data is technically accurate. A 1990 study found that after coronary bypass surgery, male patients received pain medication significantly more often than female patients, who were instead given sedatives more frequently (Calderone, 1990). An AI trained on data like this would learn to replicate that same pattern, perpetuating the very inequality it inherited.

Representation bias happens when certain groups are underrepresented in training data. An AI trained mostly on data from one demographic will simply perform worse for everyone else. A well-known example is skin cancer detection tools that are significantly less accurate for patients with darker skin tones because the training datasets contained predominantly images from fair-skinned individuals (Guo et al., 2021).

Measurement bias creeps in when the way something is measured isn’t equally accurate or fair across groups. One striking case involved an algorithm that used healthcare costs as a proxy for how sick someone actually was, and that ended up disadvantaging Black patients, who, facing disproportionate levels of poverty, tended to spend less on healthcare than equally sick white patients (Obermeyer et al., 2019).

Aggregation bias occurs when a single, one-size-fits-all model is applied to genuinely diverse populations, ignoring the fact that the same data point can mean very different things depending on a person’s background.

Learning bias emerges from technical decisions made while building the model itself, choices that can amplify disparities already present in the data, sometimes without the developers being aware. Research has shown, for instance, that differential privacy, a technique meant to protect patient confidentiality, can end up reducing a model’s accuracy for underrepresented groups even further (Bagdasaryan & Shmatikov, 2019).

Evaluation bias happens when the datasets used to test and benchmark the model don’t reflect the real population it will actually serve, meaning a system can look accurate on paper while quietly failing specific groups of patients in practice.

Finally, deployment bias arises when a tool is used in a context very different from the one it was designed and tested for, undermining its reliability in ways that are easy to miss.

Why this matters for patients

Understanding where bias comes from is the first step to catching it, and to making sure the AI tools shaping modern medicine work fairly for everyone, not just the patients who happen to be well represented in the data.

This is exactly the kind of practical, structured understanding the AEQUITAS training equips healthcare professionals with, helping them recognise these patterns before they translate into real harm for real patients.

References:

Bagdasaryan, E., & Shmatikov, V. (2019). Differential Privacy Has Disparate Impact on Model Accuracy (arXiv:1905.12101). arXiv. https://doi.org/10.48550/arXiv.1905.12101

Calderone, K. L. (1990). The influence of gender on the frequency of pain and sedative medication administered to postoperative patients. Sex Roles, 23(11), 713–725. https://doi.org/10.1007/BF00289259

Friedman, B., & Nissenbaum, H. (1996). Bias in computer systems. ACM Transactions on Information Systems, 14(3), 330–347. https://doi.org/10.1145/230538.230561

Guo, L. N., Lee, M. S., Kassamali, B., Mita, C., & Nambudiri, V. E. (2021). Bias in, bias out: Underreporting and underrepresentation of diverse skin types in machine learning research for skin cancer detection-A scoping review. Journal of the American Academy of Dermatology, S0190-9622(21)02086-7. https://doi.org/10.1016/j.jaad.2021.06.884

Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447–453. https://doi.org/10.1126/science.aax2342

Suresh, H., & Guttag, J. (2021). A Framework for Understanding Sources of Harm throughout the Machine Learning Life Cycle. Equity and Access in Algorithms, Mechanisms, and Optimization, EAAMO ’21, 1–9. https://doi.org/10.1145/3465416.3483305
AI Meets Healthcare: Understanding the Promise and the Risks

Artificial intelligence is no longer a distant, futuristic concept in medicine. It’s already here, and is becoming an everyday part of clinical practice, supporting diagnosis, clinical decision-making, treatment planning, and patient risk prediction. In fact, according to the World Health Organization, nearly three quarters of EU countries are now using AI-assisted diagnostics in some form, and that number is only expected to grow (Alowais et al., 2023; World Health Organization, 2026).

As the use of AI promises to further revolutionize the sector, what should we know before trusting it with our health?

How AI works

At its core, biomedical AI relies on three main approaches. Machine learning allows systems to learn patterns from data and make predictions without being explicitly programmed for every scenario. Deep learning, a more advanced form of machine learning, uses layered neural networks (loosely inspired by the human brain) to process complex information like medical images or clinical notes. And natural language processing (NLP) enables computers to understand and interpret human language, such as a doctor’s written notes or a patient’s description of their symptoms.

Real examples already in use

In cardiovascular care, AI tools are helping detect arrhythmias from ECG readings and estimating a patient’s individual risk of heart disease. In diabetes management, AI-powered retinal scans can catch signs of diabetic retinopathy earlier than ever before. And in mental health, natural language processing tools are being explored to help identify signs of depression in the way people write or speak, and to support patients between therapy sessions.

So what’s the catch?

Research has shown that AI applications can lead to misinterpretation and the promotion of biases and discrimination (Varona & Suarez, 2022). In essence, AI systems are only as good as the data they learn from. If that data reflects existing inequalities, for example, if it comes mostly from male patients, or predominantly from one ethnic group, the AI will inherit those biases. It will perform well for the people it “knows,” and poorly for everyone else.

Several documented cases have shown AI tools trained on biased datasets producing less accurate results for underrepresented patients, sometimes with real consequences for diagnosis and treatment.

What AEQUITAS is doing about it

This is exactly the challenge our AEQUITAS project, funded by the European Union, was designed to address. By developing training resources, a database documenting real cases of bias in biomedical AI, and a practical regulatory model for hospitals and CSOs, AEQUITAS is helping healthcare professionals and institutions across Europe use these powerful tools more safely, and more fairly, for every patient.

Understanding how AI works, and where it can go wrong, is the first step toward making sure it works for everyone.

References:

Alowais, S.A., Alghamdi, S.S., Alsuhebany, N. et al. Revolutionizing healthcare: the role of artificial intelligence in clinical practice. BMC Med Educ 23, 689 (2023). https://doi.org/10.1186/s12909-023-04698-z

Varona, D., & Suárez, J. L. (2022). Discrimination, Bias, Fairness, and Trustworthy AI. Applied Sciences, 12(12), 5826. https://doi.org/10.3390/app12125826

World Health Organization. 2026. “Artificial intelligence is reshaping health systems: state of readiness across the European Union.” World Health Organization.
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.
Fairness And Bias Mitigation in AI Models for Diabetes Diagnosis: A Comparative Evaluation of Algorithmic Approaches

AI models that predict diabetes can produce biased results for certain groups of patients, and several methods have been developed to reduce this. This study compared three of these methods on the same data, a widely used dataset of 768 women aged 21 and over, assessing fairness across age groups.

Each method had different strengths. The first relied on simple, transparent decision rules, making its predictions easy to interpret, but it had no specific mechanism to ensure fairness. The second was designed to prevent the model from basing its predictions on age, and achieved perfect results on the fairness measures used while keeping accuracy at an acceptable level. The third corrected the imbalance between patients with and without diabetes in the data, which gave the highest overall accuracy but did not ensure fairness.

The authors conclude that none of the methods achieves the right balance between accuracy, fairness and interpretability on its own, and propose combining several strategies to reduce bias in AI-based diabetes diagnosis.

Learn more about this study here: https://doi.org/10.54554/jtec.2025.17.03.004

Reference:

Mohd Hamdan, M.D.H., Ab Jabal, M.F., Abdul Rahman, S., & Kapi, A.Y. (2025). Fairness And Bias Mitigation in AI Models for Diabetes Diagnosis: A Comparative Evaluation of Algorithmic Approaches. Journal of Telecommunication, Electronic and Computer Engineering, 17(3), 27–34.
Mitigating bias in deep learning for diagnosis of coronary artery disease from myocardial perfusion SPECT images

AI can accurately diagnose coronary artery disease from myocardial perfusion imaging, a type of scan that shows blood flow to the heart muscle. However, these models are often trained on high-risk patients who were referred for invasive tests, which can lead them to overestimate the probability of disease in other patients. This study tested ways of training AI so that the probabilities it estimates match the real probability of disease more closely.

Deep learning models were trained on 828 patients from three sites and then tested on a separate group of 511 patients, including 179 with a low likelihood of disease. The best results came from a model trained with additional examples of patients without coronary artery disease and with little or no sign of reduced blood flow on the scan. It produced more accurate probability estimates than the model trained without these additional cases, and the improvement was most pronounced in women. It was also better at identifying disease than standard measures taken from the scans.

The authors conclude that adding more low-risk patients during training can improve how accurately AI estimates the probability of disease. This matters especially for lower-risk patients and for women, for whom overestimated probabilities could affect decisions about further tests and treatment.

Learn more about this study here: https://doi.org/10.1007/s00259-022-05972-w

Reference:

Miller, R.J.H., Singh, A., Otaki, Y., Tamarappoo, B.K., Kavanagh, P., Parekh, T., Hu, L.-H., Gransar, H., Sharir, T., Einstein, A.J., Fish, M.B., Ruddy, T.D., Kaufmann, P.A., Sinusas, A.J., Miller, E.J., Bateman, T.M., Dorbala, S., Di Carli, M.F., Liang, J.X., … Slomka, P.J. (2023). Mitigating bias in deep learning for diagnosis of coronary artery disease from myocardial perfusion SPECT images. European Journal of Nuclear Medicine and Molecular Imaging, 50(2), 387–397.
Evaluating and mitigating bias in machine learning models for cardiovascular disease prediction

Machine learning models are increasingly used to estimate a person’s risk of cardiovascular disease, but bias can be introduced when health data is collected and prepared, affecting how well these models work for certain groups. This study used electronic health records from 109,490 people treated at Vanderbilt University Medical Center, in the US, to test whether several machine learning models performed equally well across race and gender. It also compared them with a widely used risk equation from the American Heart Association that does not use machine learning.

Most of the machine learning models were less biased than the American Heart Association equation. Differences between racial groups were small, but differences between women and men were considerably larger. According to the authors, both the risk equation and the machine learning models were biased against women.

The researchers also tested three ways of reducing this bias. Removing race and gender from the data did not significantly reduce bias in most models, and balancing the number of people from each group in the training data did not consistently help either. Balancing the proportion of people with cardiovascular disease in each group reduced the gender bias, but not the racial bias, and slightly reduced accuracy in many cases. The authors conclude that gender disparities in cardiovascular risk prediction need to be investigated and corrected.

Learn more about this study here: https://doi.org/10.1016/j.jbi.2023.104294

Reference:

Li, F., Wu, P., Ong, H.H., Peterson, J.F., Wei, W.-Q., & Zhao, J. (2023). Evaluating and mitigating bias in machine learning models for cardiovascular disease prediction. Journal of Biomedical Informatics, 138, 104294.
Race, Sex, and Age Disparities in the Performance of ECG Deep Learning Models Predicting Heart Failure

AI models that analyse electrocardiograms (ECGs) could help identify people at high risk of heart failure early, and potentially reduce racial disparities in heart failure outcomes. But whether these models perform equally well for different groups of patients has been little studied. This study used ECGs from 326,518 patient encounters at Stanford Hospital to evaluate a model trained to predict heart failure within five years, comparing its performance by race, ethnicity, age and sex.

Overall, there were no significant differences in performance between racial groups, and performance declined with age. However, among patients aged 0 to 40, the model performed significantly worse for Black patients than for any other racial group, with the largest differences among young Black women.

The researchers tested several ways of reducing these disparities: including race, ethnicity, sex and age in the model, training separate models for each racial group, and training it on data with equal racial representation. None of these improved the disparities. Adjusting the model’s decision thresholds individually by race, age and sex did lead to substantial improvements. The authors warn that machine learning tools for heart failure risk perpetuating existing disparities, and suggest these individualised thresholds as one way to reduce them.

Learn more about this study here: https://doi.org/10.1161/CIRCHEARTFAILURE.123.010879

Reference:

Kaur, D., Hughes, J.W., Rogers, A.J., Kang, G., Narayan, S.M., Ashley, E.A., & Perez, M.V. (2024). Race, Sex, and Age Disparities in the Performance of ECG Deep Learning Models Predicting Heart Failure. Circulation: Heart Failure, 17(1), e010879.
Confounders mediate AI prediction of demographics in medical imaging

Previous research has suggested that AI can identify a patient’s age, sex and even race from medical images. This study tested whether this holds for echocardiograms, ultrasound videos of the heart, using more than 530,000 videos from two US medical centres, Cedars-Sinai and Stanford.

The models predicted sex well and age with an average error of about 9 years, and these results held across both centres. Race, however, could not be reliably predicted. When the researchers changed the balance of age or sex in the data used to train the race model, its performance varied widely, while changing the racial composition of the data had little effect on the predictions of sex and age.

The authors conclude that much of the model’s apparent ability to predict race comes from “shortcuts” through other features that differ between racial groups, rather than from race itself. They warn that treating AI predictions of race as evidence of biological differences risks reinforcing a false idea that has contributed to medical errors, since differences in health between racial groups often reflect social factors and structural racism. They call for further research into which imaging features carry demographic information, and into the risks of bias this creates in medical AI.

Learn more about this study here: https://doi.org/10.1038/s41746-022-00720-8

Reference:

Duffy, G., Clarke, S.L., Christensen, M., He, B., Yuan, N., Cheng, S., & Ouyang, D. (2022). Confounders mediate AI prediction of demographics in medical imaging. npj Digital Medicine, 5, 188.
Age, sex and race bias in automated arrhythmia detectors

Algorithms that read electrocardiograms (ECGs) automatically are increasingly used to help identify heart conditions, but few studies have examined whether they perform equally well for different groups of patients. This study compared 56 open-source algorithms, developed by participants in the 2021 PhysioNet Challenge, on more than 130,000 ECGs, grouping the results by sex, age and race.

Several performance measures were similar across groups, but others showed clear differences. Accuracy was significantly lower for Black patients and higher for Asian patients, with a gap of 39% on the test set. Accuracy also fell with age, by around 12% per decade, and results for male and female patients differed depending on the measure used.

The researchers then retrained one of the best-performing models with an additional constraint designed to reduce performance differences across sex, race and age. This led to a small loss in overall performance but a significant reduction in bias. The authors conclude that bias depends on the model’s design, the population it is trained on and how its performance is measured, and that all of these should be closely examined.

Learn more about this study here: https://doi.org/10.1016/j.jelectrocard.2022.07.007

Reference:

Perez Alday, E.A., Rad, A.B., Reyna, M.A., Sadr, N., Gu, A., Li, Q., Dumitru, M., Xue, J., Albert, D., Sameni, R., & Clifford, G.D. (2022). Age, sex and race bias in automated arrhythmia detectors. Journal of Electrocardiology, 74, 5–9.
Artificial intelligence-enhanced electrocardiography for accurate diagnosis and management of cardiovascular diseases

Electrocardiography (ECG) enhanced by artificial intelligence is emerging as a promising tool for diagnosing and managing cardiovascular disease. This review explains how AI can detect subtle patterns in the ECG that human interpreters may miss, helping to identify problems such as arrhythmias and heart failure more quickly and consistently.

However, the authors also highlight important limitations, including systematic bias. When training datasets do not adequately represent different ages, sexes and racial groups, models can perform worse for those who are underrepresented, since ECG patterns vary with age, physiological differences between the sexes and genetic factors. To make these tools safe and fair in clinical practice, the review calls for more diverse datasets, bias mitigation strategies and continuous validation across populations.

Learn more about this review here: https://doi.org/10.1016/j.jelectrocard.2024.01.006

Reference:

Muzammil, M.A., Javid, S., Afridi, A.K., Siddineni, R., Shahabi, M., Haseeb, M., Fariha, F., Kumar, S., Zaveri, S., & Nashwan, A.J. (2024). Artificial intelligence-enhanced electrocardiography for accurate diagnosis and management of cardiovascular diseases. Journal of Electrocardiology, 83, 30–4
Cardiovascular Disease Screening in Women: Leveraging Artificial Intelligence and Digital Tools

Cardiovascular disease remains the leading cause of death in women, and there is growing evidence of sex- and gender-based differences in how it develops and progresses. This review provides an overview of the main opportunities for cardiovascular screening in women and of the potential role of artificial intelligence and digital tools.

The authors note that healthcare providers, guideline committees and policymakers are not yet well equipped to collect data on women, interpret it, or use it to design targeted interventions. The review includes a section on bias in AI and looks at different stages of a woman’s life, from youth and pregnancy to menopause and older age, as well as the implications of AI for clinical guidelines and public health.

Learn more about this review here: https://doi.org/10.1161/CIRCRESAHA.121.319876

References:

Adedinsewo, D.A., Pollak, A.W., Phillips, S.D., Smith, T.L., Svatikova, A., Hayes, S.N., Mulvagh, S.L., Norris, C., Roger, V.L., Noseworthy, P.A., Yao, X., & Carter, R.E. (2022). Cardiovascular Disease Screening in Women: Leveraging Artificial Intelligence and Digital Tools. Circulation Research, 130(4), 673–690.
Implicit gender bias in the diagnosis and treatment of type 2 diabetes: A randomized online study

This study examined whether implicit gender biases, unconscious associations about people based on their gender, influence how general practitioners (GPs) diagnose and treat type 2 diabetes. Ninety-nine GPs completed tests measuring two biases, one related to lifestyle (women having a healthier lifestyle than men) and one related to communication (men being less communicative than women), and then assessed clinical cases featuring a male or female patient.

Female GPs showed a significant lifestyle-related bias, while GPs of both genders showed a communication-related bias. The patient’s gender also affected clinical decisions: for example, GPs were less certain about the diabetes diagnosis when the patient was a woman.

These findings show how unconscious bias can shape clinical decision-making.

Learn more about this study here: https://doi.org/10.1111/dme.15087

Reference:

Skvortsova, A., Meeuwis, S.H., Vos, R.C., Vos, H.M.M., van Middendorp, H., Veldhuijzen, D.S., & Evers, A.W.M. (2023). Implicit gender bias in the diagnosis and treatment of type 2 diabetes: A randomized online study. Diabetic Medicine, 40(8), e15087
AEQUITAS
Proposal: Preparation of the CSOs and public healthcare sector to address gender and racial biases that might arise from the wide usage of AI in order to protect and promote fundamental rights

Implementation: 2025 to 2027

Call: CERV-2024-CHAR-LITI – Promote civil society organizations’ awareness of, capacity building and implementation of the EU Charter of Fundamental Rights

Proposed Budget: 1 061 761,00€

Keywords: gender and racial biases, biomedical AI, EU Charter of Fundamental Rights

Objective: The topic of the AEQUITAS project is the gender and racial biases that have been reported in biomedical AI and which can lead to misdiagnosis and mistreatment and how they pose a threat to the fundamental rights protected by the EU Charter. The aim of the project is 3-fold:

– to increase the capacity of the CSOs and human rights organizations in educating the public, monitoring the biases and advocating for the protection of the fundamental rights especially regarding the biomedical AI;

– to increase the knowledge of healthcare staff from public hospitals on the biases that biomedical AI can present due to the biased data from which they were fed and to help them approach and consult these AI systems with a critical mindset;

– to develop an AI Regulatory Model that will be used by CSOs and public hospitals in their practices;

– to develop a European network of CSOs and public hospitals that will collaborate and support each other in the effort to raise public’s awareness on the gender and racial biases of biomedical AI and on the applications of the EU Charter;

– to raise awareness of the EU Charter of fundamental rights and its application in the AI era;

– to develop and distribute policy recommendations in order to advocate for the need to regulate biomedical AI.

Partners:
- Innovation Hive – Kypseli Kainotomias
- Kentro Ginaikeion Meleton Kai Ereyvnon Astiki Mi K
- Universitat Zu Koln
- Center for the Study of Democracy
- C.I.P. Citizens in Power
- Moterų Informacijos Centras Asociacija Mic
- Lobby Europeo de Mujeres en Espana LEM España
- TIA Formazione Internazionale Associazione APS
- Health Citizens – European Research Institute
- Technologiko Panepistimio Kyprou
- Cyens Centre of Excellence
- Edex – Educational Excellence Corporation Limited
- Rite Research Institute for Technological Evolution
- Vucable
- TotalEU Production
Project Website: https://aequitasproject.eu/