AI could shorten endometriosis diagnosis by spotting symptom patterns in health records, reading ultrasound and MRI more consistently, and powering new non-invasive tests. But models learn from past diagnoses, and past diagnoses reflect years of dismissed pain. Without deliberate design, AI may copy the delay it was meant to fix.
A teenager misses school two days a month because of pain. Her GP suggests painkillers and the pill. At 19 she is told it is probably irritable bowel syndrome. At 23 a gynaecologist orders an ultrasound that shows nothing. At 27, trying to conceive, she finally has a laparoscopy, and a surgeon sees what was there all along.
Versions of that story are so common that they have become the defining feature of endometriosis. The World Health Organization estimates that the condition affects about 10% of women and girls of reproductive age, around 190 million people, and that the average time to diagnosis ranges between 4 and 12 years. The WHO also points to a root cause that is not technical at all: the public, families and most health workers do not recognise that the chronic pelvic pain of endometriosis is not normal.
Artificial intelligence is being offered as a way to cut through this. The thesis here is that the diagnostic delay is partly a data problem that AI can genuinely help with, and partly a bias problem that AI is very well placed to reproduce. Which one wins depends on how the tools are built and where they are placed in the care pathway.
Why does endometriosis take so long to diagnose?
The delay has several layers, and it helps to separate them.
The first is patient-level normalisation. Many girls and women are told, or come to believe, that severe period pain is part of life. They present late or not at all.
The second is clinical. Symptoms overlap with irritable bowel syndrome, pelvic inflammatory disease, urinary problems and mental health conditions. Clinicians see fragments of the picture across separate visits, often to different specialists, and rarely join them up. Pain is subjective and easy to discount.
The third is diagnostic technology. For decades the reference standard was laparoscopy, surgery to look inside the pelvis. That set a high threshold for confirming the diagnosis. Imaging has improved, and the WHO now describes clinical diagnosis based on symptoms and imaging, without surgery, as a legitimate route. But transvaginal ultrasound and MRI for endometriosis require specialist skill, and superficial disease, the most common form, can be invisible on imaging even in expert hands.
The fourth is structural. Gynaecology waiting lists are long in many countries, specialist endometriosis centres are few, and women’s health conditions have historically received less research funding relative to their burden.
Each layer suggests a different role for AI, and a different risk.
How can AI help diagnose endometriosis earlier?
There are three main approaches, at very different stages of maturity.
Pattern-finding in health records. The typical patient’s record contains years of scattered clues: repeated consultations for pelvic pain, painkiller prescriptions, visits for bowel or bladder symptoms, fertility concerns. A model that reads the whole record could flag women whose pattern resembles that of people later diagnosed, prompting earlier referral. This addresses the second layer, fragmented care, directly. Several research groups have built such models retrospectively. The evidence that they change time to diagnosis in practice is still thin.
Imaging support. AI tools to help detect deep endometriosis and endometriomas on ultrasound and MRI aim to make specialist-level interpretation more widely available. This addresses the third layer. But imaging AI cannot see what imaging cannot show, and superficial peritoneal disease often leaves no trace on a scan.
Non-invasive biomarker tests. The most commercially advanced example is a saliva test based on microRNA signatures analysed with machine learning, developed in France and marketed as the Endotest. Its developers have published studies reporting high accuracy, and a developer-led cost-effectiveness analysis appeared in BJOG in 2023. The evidence base illustrates both the promise and the pitfalls. One 2024 study focused on superficial peritoneal endometriosis reported perfect sensitivity and specificity for its best model, but that subgroup included only 16 cases. A result like that says more about the size of the sample than about how the test will perform in thousands of women in general practice.
That is not a criticism unique to one test. It is the standard problem of diagnostic AI: models built and validated in specialist centres, on patients who have already been referred and who mostly have the disease, tend to look far better than they will among the broader population in primary care, where most women with pelvic pain do not have endometriosis. Independent, prospective validation in the setting of intended use is what matters.
How could AI copy the bias in endometriosis diagnosis?
This is the part of the story that deserves more attention than it gets.
A model trained to predict endometriosis learns from people who were diagnosed. But who gets diagnosed is itself the product of the delay. Women who were taken seriously, had access to specialists, could afford private care or pushed hard for referral are overrepresented among confirmed cases. Women whose pain was dismissed, including many from minority ethnic groups, poorer communities and places without specialist centres, are underrepresented, or appear in the data as “undiagnosed pelvic pain” labelled as something else.
A model trained on that history will learn that certain profiles are more likely to have endometriosis, and it may be learning who gets referred rather than who has the disease. The historical myth that endometriosis mainly affects white, affluent, career-focused women is a warning here: it shaped who was investigated for decades, and data generated under that belief still sit in health records.
There are other ways the bias enters. If a symptom checker uses language patterns, women who describe pain differently, or in a second language, may score lower. If the model uses prior specialist visits as a feature, it will favour those who already had access. And if it is deployed only in well-resourced clinics, it will widen the gap between those who are already better served and those who are not.
We examined this general problem in our piece on bias in clinical algorithms: the most effective fix is often checking the label, what the model is being trained to predict, rather than tweaking the model. For endometriosis, the label “diagnosed” is contaminated by the very delay we want to reduce.
A worked example: two ways to build the same tool
Hypothetical, for illustration. A health system wants a primary care tool that flags women with pelvic pain for specialist referral.
Version one trains on records of women with a confirmed diagnosis versus everyone else. It performs well on retrospective data. In deployment, it flags mostly women with long histories of specialist visits and private care, because those features predict a confirmed diagnosis. Women with years of GP visits for pain but no referral are scored as lower risk, because in the training data such women rarely received a diagnosis.
Version two defines the target differently: persistent cyclical pelvic pain meeting a clinical threshold, regardless of whether a specialist confirmed it. It deliberately excludes features that reflect access, such as private care or prior specialist visits. It is validated separately across ethnic groups, age bands and deprivation levels. Its retrospective accuracy for predicting a confirmed diagnosis looks worse, because it flags women the old system never confirmed. In practice, it surfaces precisely the women the delay has been hiding.
The second version is harder to sell, because its headline accuracy is lower. It is also the one more likely to reduce the delay.
What is the strongest case for relying on AI here?
The strongest counterargument is that the status quo is already so poor that even a biased tool would help. If the average woman waits years, a model that flags some of them earlier, even imperfectly, reduces suffering. Waiting for a perfect, unbiased tool means more years of delay.
There is force in this. The bar set by current practice is low. But a tool that helps mainly those already better served will be counted as a success by aggregate measures while leaving the worst-served women where they were, or further behind as specialist capacity is absorbed by the flagged group. The appropriate response is not to wait for perfection but to measure performance and benefit by subgroup from the start, and to set the target label with care.
A second counterargument is that the real bottleneck is specialist capacity, not detection. Flagging more women faster is pointless if gynaecology waiting lists stretch for a year. That is true in many systems, and it argues for AI tools that help primary care start treatment and manage symptoms, not only for tools that generate referrals.
What this means for health systems, clinicians and developers
For health systems, treat AI for endometriosis as part of a pathway redesign, with clear primary care management, referral criteria and specialist capacity planned together. Require developers to show performance by ethnicity, age and deprivation, and monitor who is flagged after deployment.
For GPs and gynaecologists, a low AI score should never override a clinical picture of persistent, cyclical pelvic pain. Many women can start effective treatment for suspected endometriosis without a confirmed diagnosis, following guidelines. Listen to the history.
For developers, define the target label thoughtfully, exclude features that encode access to care, validate prospectively in primary care populations, and publish subgroup results. Be cautious about headline accuracy from small or specialist cohorts; our analysis of where AI diagnostic evidence is strongest explains why narrow, well-defined tasks validate best.
For regulators and payers, reimbursement decisions for new diagnostic tests should require independent validation in the population of intended use, not only developer-led studies.
For women with pelvic pain, if period pain stops you doing normal activities, or pain persists despite treatment, it is reasonable to ask your doctor directly whether endometriosis could be the cause and what the next steps are.
Frequently asked questions
Can AI diagnose endometriosis?
AI tools are being developed to flag likely endometriosis from health records, support ultrasound and MRI interpretation, and analyse biomarkers such as salivary microRNA. Some report high accuracy in specialist cohorts, but independent validation in primary care is limited. AI may support earlier referral, but diagnosis still rests on clinical assessment, imaging and sometimes surgery.
How long does it take to diagnose endometriosis?
The World Health Organization reports that the average time to diagnosis ranges between 4 and 12 years. Delays arise from normalisation of period pain, symptoms overlapping with other conditions, limited specialist imaging skills and long waiting lists. Endometriosis affects an estimated 10% of reproductive-age women and girls, around 190 million people worldwide.
Is there a saliva test for endometriosis?
A saliva test based on microRNA analysed with machine learning has been developed in France and marketed as the Endotest. Developer-led studies report high accuracy, but some analyses involved small subgroups, and broader independent validation is still needed. Ask your specialist whether such a test is available and appropriate for you.
Can AI be biased in diagnosing endometriosis?
Yes. Models learn from past diagnoses, and past diagnoses favoured women with better access to specialists. If a model learns who was referred rather than who had the disease, it may underrate women whose pain was historically dismissed. Checking what the model predicts, and testing performance across groups, reduces this risk.
The endometriosis delay was built by people not believing women’s pain. AI can help undo it, but only if the people building it are careful not to train a machine on the same disbelief.
If you have lived with endometriosis or treat it, where in the path to diagnosis would a tool have made the most difference, and what would you worry about?
