Why we need more women’s health data for AI models

As AI technology becomes more integrated into healthcare, the data on which these models are trained must accurately represent the populations they serve. Unfortunately, existing datasets often lack sufficient representation of women, leading to models that may not perform effectively. While there are some large publicly available women’s health datasets, there needs to be increased collaboration between industry, research, and federal agencies to effectively leverage AI to advance women’s health care.  

How did we get here?

There is a long history of women being excluded from clinical research. In 1977, a Food and Drug Administration policy recommended excluding women of childbearing potential from Phase I and early Phase II drug trials. This oversight was rooted in the belief that men and women were biologically similar enough that results from male studies could be generalized to women, and that a woman’s hormones could confound research findings. There were also concerns with pregnancy risk and fetal abnormalities as an unintended consequence from pharmaceutical trials even when women were on reliable forms of birth control. The Thalidomide scandal, which led to more than 10,000 children born with severe physical deformities, further propagated the exclusion of women from clinical research.  

However, this perspective was challenged in subsequent decades. The National Institutes of Health (NIH) recognized the necessity of including women in research in the early 1990s. In 1993, the NIH Revitalization Act mandated that women and minorities be included in clinical trials, setting a precedent for better representation in health studies. This legislation not only aimed to rectify past biases but also encouraged researchers to stratify results by gender, paving the way for a deeper understanding of sex differences in health and disease. Despite these federal policies, various studies of NIH funding estimate that only 30% of funded research include data on sex or gender as a variable.  

This has real consequences. Failure to study medications and other interventions has led to women experiencing adverse side effects from medications at twice the rate of men. Additionally, “diseases that predominantly affect women — such as migraines, headaches, anorexia and endometriosis — received funding that was a fraction of what was awarded for diseases that predominantly affected men, when funding amounts are matched to disease burden,” according to a 2023 article in Nature. This speaks to limitations in the amount of data available to train AI models to address common women’s health issues.  

Women’s health data: What’s available and what’s lacking

The need for more comprehensive women’s health data is pressing, especially as we venture into an era of artificial intelligence in healthcare. AI models are fundamentally reliant on the quality and diversity of the data they are trained on. High-quality data that is accurate, relevant, and representative will lead to better-performing models. AI models that lack high-quality data can inadvertently perpetuate or even exacerbate existing biases present in the training data. Large datasets often provide more variability, which can enhance the learning process of the model. However, it’s important to note that more data is not always better if it includes a significant amount of irrelevant information. Quality remains paramount. 

There are a number of research datasets with women’s health data available, which could be used for AI models as a starting place. The All of Us Research Program is an NIH effort to gather health information from one million or more people living in the United States. This dataset currently includes over half a million participants, 60% which are women, and consists of medical records, genomics, wearable, and health survey data. The Nurses’ Health Study (NHS) is a long-term study of women’s health, initiated in 1976, which aims to investigate the risk factors for chronic diseases in women. This dataset includes self-reported questionnaire data and biospecimens, however the criteria for participation was married female registered nurses between the ages of 30-55 which automatically excludes those either earlier in their reproductive years or postmenopausal. The Women’s Health Initiative (WHI) is a comprehensive study initiated in the 1990s to address major causes of morbidity and mortality in postmenopausal women. The dataset includes information on a wide range of health issues, such as cardiovascular disease, cancer, and osteoporosis.  

It’s important to note that all of these are observational studies, which gather information and look for correlations, versus interventional studies, such as pharmaceutical trials which examine cause and effect of exposures. These datasets are limited in scope, often focusing on specific conditions or demographics, leading to slices of information and lacking robust, diverse, longitudinal, data that can easily be combined. To develop AI models that can accurately assess and address women’s health needs, it is crucial to expand and enhance these datasets. 

A way forward

To improve AI model development for women’s health, several strategies should be employed. First, encouraging public-private partnerships could facilitate the sharing of anonymized health data while ensuring patient privacy. For example, pharmaceutical companies possess enormous vaults of proprietary data collected during clinical trials. Combining this data with NIH-funded study data and de-identified medical records data could lay the groundwork for robust AI models to train on.  

Second, companies like Clue, Flo, Oura, and Apple all collect menstrual cycle data from their users. Clue’s users can opt-in to contribute their de-identified data for research purposes. The Apple Women’s Health Study in partnership with Harvard T.H. Chan School of Public Health and NIEHS aims to advance the understanding of menstrual cycles and how they relate to various health conditions such as polycystic ovary syndrome (PCOS), infertility and menopausal transition through user cycle tracking. Combining either passively collected data from wearable devices or user-reported menstruation duration and symptom data could contribute to an increasingly robust dataset for AI to understand women’s reproduction patterns and the menstrual cycle as a vital sign. 

By implementing these strategies, we can pave the way for innovative AI applications in healthcare that truly serve all genders equitably and improve health outcomes for women. Then maybe we can start to reduce the 7 year timeframe for diagnosing endometriosis, understand why 78% of Americans with autoimmune diseases are women, why women are 50% more likely to die in the year following a heart attack than men, or why Alzheimer’s affects women twice as often as men. The answers could be in the big data.  

Katie D. McMillan, MPH is the CEO of Well Made Health, LLC, a business strategy consulting firm for health technology companies. She is also a curious researcher and writer focusing on digital health evidence, healthcare innovation, and women’s health. Katie can be reached at katie@wellmadehealth.com or LinkedIn.