Author:
Artificial intelligence (AI) is rapidly becoming an integral tool in women’s health, from fertility to pregnancy to menopause care. It has the potential to identify health risks earlier, personalize care, and connect people to the right providers at the right time. But there is a blind spot. AI can only learn from the data we give it. And in women’s health, that data is incomplete. AI has the power to transform women’s health. But realizing that potential requires more complete data and the human clinical expertise to interpret it in the context of women's individual needs and experiences.
Without addressing the data gaps, we risk teaching powerful new technology to reproduce longstanding bias at scale. To understand what’s at stake, consider how incomplete evidence has already shaped women’s healthcare.
Where Healthcare Data Misses Women’s Needs
A recent study of more than 2,400 women with clinically significant menopause symptoms found that 63% had sought care, yet only 23% had a menopause-related diagnosis code in their medical records. In other words, women were seeking help, but their medical records often failed to reflect their needs. If AI learns only from what is documented, it risks overlooking the very patients it should be helping. The absence of a diagnosis code does not mean the absence of a clinical need. (source)
Pregnancy may seem easy for AI to identify, but the data tells a different story. Full-term births are well documented, while early pregnancy, miscarriage, ectopic pregnancy, and preterm birth can be harder to capture accurately. One study found claims and electronic health record data agreed on 97.8% of full-term births, but only 62.4% of preterm births. Natural language processing can fill gaps by extracting information from clinical notes, but it still depends on what clinicians document and how consistently they document it. Better technology can make better use of available data, but it cannot create a complete picture from incomplete information. (source)
Evidence Needs Clinical Context
The consequences of incomplete or misinterpreted evidence predate AI. Consider the widely cited 2002 Women’s Health Initiative (WHI) study, which didn’t tell the whole story about the risks and benefits of hormone therapy (HRT). The study found that women taking the specific hormone formulations studied had increased risk of breast cancer and cardiovascular complications than those taking the placebo. Its findings, based largely on a population of women with an average age of 63, were widely generalized to HRT overall, creating the perception that hormone replacement therapy was dangerous. It obscured important differences in age, timing, formulations, and individual risk (source). As the evidence was clarified and awareness of appropriate HRT use grew, demand surged, contributing to shortages of HRT products.
Our own experience shows how healthcare treatment and access can change when the data paints a more accurate picture, and people receive that information in a clear, relevant way. At WIN, our menopause consultations increased by 15x in the two months following the FDA removal of its black box warnings about cardiovascular disease, breast cancer, and dementia from menopausal hormone therapies. The evolution of the HRT conversation shows what happens when evidence is incomplete or misunderstood. The lesson for AI is clear: without clinical context, findings about one population can become assumptions about everyone. On the flip side, when people receive relevant, credible information, they are more likely to seek care. AI’s promise depends on getting both right: interpreting evidence in context and helping people act on information that is relevant to their care.
Low Engagement Can Hide Unmet Needs
The data gap also matters to employers and their benefits programs. Employers may offer fertility, maternity, menopause and other women’s health benefits to their entire workforce. But only a portion of employees will actively engage with these programs. The employees who do engage generate more data. They may have a fertility assessment, talk to a nurse advocate, schedule an appointment with an OB/GYN or reproductive endocrinologist, track data with a wearable, or use an app. These interactions make their health needs more visible to healthcare systems and AI models. People who may be struggling with any of these areas of health but do not engage generate fewer signals in the data. If those signals guide outreach and support, employees already engaged in care may receive more attention, while those with fewer recorded interactions remain overlooked. The risk is that employers and AI models mistake the needs of those who engage for the needs of the entire workforce, widening the gap.
Better AI Requires Human Clinical Judgement
Closing these gaps requires us to look beyond what is easiest to measure. Claims and medical records are valuable, but they need to be supplemented by what women report about their symptoms, experiences and barriers to care. Clinical expertise and human conversations provide context that a diagnosis code alone cannot. That context must shape how we build, train, and evaluate AI, including whose needs it recognizes and whose it continues to miss. We must also evaluate whether these tools recognize needs among people with limited records or little prior engagement. Keeping clinicians involved in interpreting AI-generated insights helps ensure that decisions reflect not only what the data reveals, but also the individual circumstances, needs, and experiences data alone may miss.
The opportunity in front of us is to pair responsible AI with clinical expertise and human support to identify unmet needs earlier, understand barriers to engagement, and connect people with appropriate care. AI can help identify patterns and guide people toward resources, but when healthcare decisions are complex, sensitive, and deeply personal, clinical expertise and a trusted human support remain essential. The risk is that we’ll get better at serving people who already know how to navigate healthcare while leaving behind those who face the greatest barriers. Success should be measured by whether more women receive appropriate support and better outcomes. AI should not simply make care and intake more efficient. It should make high-quality care easier to access, understand and use. We need to use it to close the gap, not widen it.
