Women receive less active medical treatment than men, raising concerns for healthcare AI

Women with the same medical condition as men appear less likely to be offered active medical management, including surgery, a stent or stronger pain relief, according to research from the University of St. Andrews. The findings raise a difficult question for healthcare providers. Are differences in treatment supported by relevant clinical factors, or are established patterns of medical practice resulting in women receiving less active care? They also present a technological challenge. If artificial intelligence systems are trained on medical literature and clinical records containing unexplained disparities, these systems could reproduce the same patterns across future treatment recommendations.

The St Andrews researchers reviewed 1,112 published studies before identifying those that directly compared care given to male and female patients. Among 38 studies that analysed patient records, 33 reported a significant difference in the treatment received. The pattern appeared across cardiology, surgery, transplantation and emergency medicine. In most of the studies, the treatment difference remained after statistical adjustment for other relevant variables. The research was published in the peer-reviewed journal PLOS ONE.

The review does not demonstrate that every difference in treatment was discriminatory or clinically inappropriate. Treatment decisions can legitimately be affected by a patient’s medical history, risk factors, co-existing conditions, treatment preferences and likely response to an intervention. However, the St Andrews researchers found that almost none of the studies referred to a clinical guideline recommending different treatment according to sex. This leaves unresolved whether the observed differences resulted from appropriate individual clinical judgement, differences not adequately captured in the studies, or unequal care.

“For clinicians, the findings are a prompt to check whether treatment is being offered on clinical grounds rather than assumption,” study co-lead Dr Andrew O’Malley said in the material accompanying the research, which was sent to Digital Journal.

O’Malley pointed to a study of decisions about advanced heart-failure therapy. When the clinical teams responsible for selecting patients functioned poorly, women were less likely to be chosen. He also noted evidence that doctors can be more likely to attribute women’s symptoms to anxiety and can make diagnostic errors involving female patients even when test results are positive. When a difference concerns access to pain relief, surgery, transplantation or cardiovascular intervention, the potential consequences are substantial.

Medicine’s historical evidence gap?

Dr Miriam Veenhuizen, an Honorary Lecturer in the University’s School of Medicine, said the direction of the findings was not unexpected, but their consistency was striking. “The same pattern appeared in cardiology, surgery, transplant medicine and emergency care, and it survived statistical adjustment in most studies,” she said.

One possible explanation is the historical under-representation of women in clinical trials. If clinical guidelines and medical education are based disproportionately on evidence obtained from men, clinicians may have less reliable information about disease presentation, treatment responses and risks in women. This does not mean that female and male patients should invariably receive identical treatment. Biological sex can affect disease risk, symptoms, pharmacology and treatment response. The requirement is for any difference in care to rest on robust evidence and the patient’s clinical circumstances, rather than assumptions about sex or gender.

The researchers also identified a significant imbalance in academic attention. Veenhuizen said that 551 studies published during the period examined inequality affecting doctors and other health professionals, while only 41 investigated what happened to patients. This patient-level evidence gap is consequential because historical clinical decisions are increasingly becoming training data for digital systems.

A Canadian parallel

The issue has particular relevance for Canada. The Canadian Institutes of Health Research’s National Women’s Health Research Initiative was created specifically to address under-researched and high-priority areas of women’s health and to turn evidence into improved care and health outcomes. The initiative applies an intersectional approach, including attention to Indigenous, Black and racialised women, women with disabilities and gender-diverse communities. The federal government committed C$20 million over five years, from 2022 to 2027, to the initiative. Contributions from several CIHR institutes increased the total investment to C$26.9 million. CIHR describes it as a first-in-a-generation federal investment in women’s health research in Canada.

The broader CIHR Institute of Gender and Health has a mandate to investigate how sex and gender influence health and to apply that knowledge to health challenges affecting women, men and gender-diverse people. Canadian cardiovascular evidence illustrates why this work is necessary. The Heart and Stroke Foundation of Canada reports that two-thirds of heart disease and stroke clinical research has focused on men. It also states that the symptoms of half of women who experience a heart attack go unrecognised.

The foundation identifies gaps in research, diagnosis and care as threats to women’s heart and brain health. Its strategy includes expanding women-focused research, improving public and professional knowledge, developing better models of care and addressing systemic barriers. These Canadian findings do not prove that the precise pattern identified at St Andrews occurs uniformly across every province, speciality or clinical setting. They do, however, demonstrate that Canada recognises significant sex- and gender-related deficiencies in medical evidence, diagnosis and care.

The St Andrews researchers now intend to test whether generative AI systems produce similar treatment differences. This is important because AI learns statistical relationships from existing material. In healthcare, the relevant sources can include research papers, medical guidelines, electronic records and examples of past clinical decisions. Historical data records valuable medical knowledge, but it may also encode who was studied, whose symptoms were recognised and which patients received more active treatment. An algorithm does not need an explicit instruction to treat women differently. It can learn that pattern from the decisions represented in its training data.

Canada’s digital-health policies increase the urgency of examining this risk. Under federal-provincial health agreements, governments have committed to improving how health information is collected, shared, used and reported. This includes securely sharing comparable information, adopting common interoperability standards and applying person-centred principles to health-data management. Canada is also investing heavily in AI research, commercialisation and standards through its Pan-Canadian Artificial Intelligence Strategy. The programme supports the national AI institutes Amii, Mila and the Vector Institute, while the Standards Council of Canada is responsible for advancing AI-related standards.

This combination creates both an opportunity and a risk. Larger and better-connected datasets could help researchers identify inequalities that would be difficult to see in individual consultations. Conversely, an AI model trained without careful attention to sex, gender and intersecting patient characteristics could automate old disparities.

The overall insight is that AI systems used in healthcare should therefore be assessed not only for average accuracy, but also for how they perform across clinically relevant patient groups. Testing should examine whether comparable cases receive comparable recommendations, whether differences can be clinically explained and whether model performance changes across sex, gender, ethnicity, age, disability and geography. Canadian validation is particularly important because models developed elsewhere may not represent the country’s population or provincial healthcare systems.