The Diagnosis Gap: Why Women's Health Needs Better Data Before Better Algorithms

Apr 2026

I have been spending a lot of time in medical journals lately. As someone building in the women's health space, I wanted to understand where the research actually stands, what AI can do today for conditions like PCOS and endometriosis, and more importantly, where it falls short. What I found was both encouraging and frustrating in equal measure, and I want to share some of what I have learned.

The starting point is a number that has barely moved in decades. According to a systematic literature review published in BJOG, the average time to diagnose endometriosis ranges from 5.4 to 11.4 years, with patients seeing five or more doctors before receiving a diagnosis (De Corte et al., 2024). The World Health Organization reports a similar range of 4 to 12 years (WHO, 2023). PCOS follows a parallel pattern. These are not rare conditions. Endometriosis affects 5 to 10 percent of people who menstruate, and PCOS affects 6 to 13 percent. In that diagnostic window, women watch their fertility, mood, and quality of life erode while they wait for someone to listen.

Recent research suggests AI could help close this gap. A study in Frontiers in Endocrinology used machine learning models trained on electronic health records, including menstrual cycle length, BMI, and hormonal markers, to predict PCOS diagnosis, achieving an AUC of 85 percent (Barrera et al., 2024). A study in Scientific Reports found that combining clinical and ultrasound features enabled highly accurate, non-invasive PCOS detection, with follicle count, weight changes, AMH levels, and menstrual irregularity emerging as the most predictive features (Ahmed et al., 2025). On the endometriosis side, a machine learning model trained on self-reported symptom questionnaires achieved a sensitivity of 93 percent and specificity of 95 percent, demonstrating that the signal is there in the patterns of symptoms women experience daily (Sivajohan et al., 2023). A Bayesian network study further showed that specific pain locations and pain types can forecast an endometriosis diagnosis (Jeu et al., 2024).

The research is promising. But here is the part that keeps me up at night as an engineer.

Machine learning models are only as good as the data they are trained on. And women's health data is thin. The male body has historically been treated as the default in clinical research (AAMC, 2024). As recently as 2019, women accounted for roughly 40 percent of participants in clinical trials for cancer, cardiovascular disease, and psychiatric disorders, despite representing 51 percent of the population (Labiotech, 2025). Only 7 percent of biopharma R&D is invested in conditions that exclusively affect women (DIA Global Forum, 2025). Conditions biologically unique to us, like endometriosis, PCOS, and menopause, lack the robust, representative datasets that modern AI requires to generalize. A review in Women's Health put it plainly: existing AI models for endometriosis are largely proof-of-concept, limited by small sample sizes and narrow populations (Dungate et al., 2024).

When you train a model on a narrow population, it performs well in controlled settings and fails quietly in the real world. It misses the woman whose symptoms present differently. It misdiagnoses. Research has shown that neural networks trained on large medical imaging datasets underdetect disease in female patients (Cirillo et al., 2024). In early 2025, the UK Minister of State for Women's Health warned that without representative data, AI could entrench disparities rather than reduce them (World Economic Forum, 2025). The tools we build to help women could end up failing them in new, harder-to-detect ways.

So what do we actually need?

This is something I think about not just as a builder, but as a student. In my graduate work at Harvard, I study the foundations of what makes machine learning models actually work, and so much of it comes back to data. We learn early that a model's ability to generalize depends entirely on whether the training distribution reflects the population it will serve. When it does not, you get distribution shift, and the model fails silently on the patients it has never seen. We learn that evaluation metrics like accuracy or AUC can be deeply misleading when the test set is not representative, painting a picture of performance that falls apart in deployment. And we spend significant time on bias mitigation and data preprocessing, because the truth is that no amount of architectural sophistication can compensate for a dataset that does not represent the people it is meant to help. These are not abstract lessons. They describe exactly what is happening in women's health AI today.

We need better data before we need better algorithms. We need longitudinal symptom tracking at scale. Not just clinical trial data, but the lived, daily experience of women tracking their cycles, logging their symptoms, noting what feels off. A review in Frontiers in Digital Health noted that wearable devices could provide granular, objective data for endometriosis, yet to date, almost no studies have utilized wearable technology to track endometriosis symptom trajectories (Gater et al., 2023). We need diverse cohorts that reflect the full spectrum of how these conditions present across age, race, and body type. We need to treat patient-generated data not as noise, but as signal.

This is part of why I am building what I am building. Reen is not just an app. It is an attempt to create the dataset that should already exist. Every symptom logged, every cycle tracked, every pattern surfaced is a small contribution to a body of evidence that has been neglected for too long.

The AI is promising. The research is encouraging. But we are not there yet, and we will not get there by training models on incomplete pictures of women's health. The six-to-ten-year diagnosis gap is not just a medical failure. It is a data failure. And closing it will require engineers, researchers, and builders who understand that the most important thing we can do right now is not build a smarter model. It is build a more complete picture of what women's health actually looks like.

References