Decades of male‑focused medical research could bias health care AI

Decades of medical research that has focused primarily on male patients may be influencing the development of health care artificial intelligence (AI). This is because much of the research and training data used to train AI systems has been based on medical records and studies that have often treated the male body as the norm. For example, a 2024 study found that three-quarters of CPR manikins sold worldwide were described as male or did not specify sex, with only one offering a breast overlay.

Similarly, out-of-hospital cardiac arrest statistics show that women are less likely to receive bystander CPR compared to men, and they have a higher risk of not surviving to hospital discharge if CPR is provided more than four to five minutes after a witnessed cardiac arrest. Furthermore, studies have shown that women and men with similar symptoms have historically been assessed differently, which could lead AI algorithms to learn and perpetuate these patterns.

Health care AI also learns associations from health data, which may reflect clinical decisions as well as biological factors. If women and men have historically been assessed differently due to their sex, an AI system trained on such data might interpret these patterns as evidence of disease rather than a reflection of healthcare disparities.

This issue is compounded by the fact that women are consistently underrepresented in medical studies, and hormonal changes during the menstrual cycle have often been viewed as unwanted variation. Regulatory policies and research practices have historically favored male subjects and animals, despite evidence that females are biologically no more variable than males.

However, there have been improvements in representation, with the 2016 Sex and Gender Equity in Research (SAGER) guidelines encouraging researchers to consider and report sex and gender. Nevertheless, equal participation does not always translate to equal analysis, as only 44% of 574 reviewed papers analyzed results by sex. Given that medical research is increasingly becoming more representative, it is crucial for AI developers to ensure that their systems are trained on balanced and representative datasets to avoid embedding historical biases in new technologies.