Women spend around a quarter more of their lives in poor health than men, and for many conditions the path to a diagnosis still takes years. One reason is a data problem: women remain underrepresented in clinical evidence, and the symptom patterns that matter most are often the hardest to describe in a short appointment. With the SympATA app, Fraunhofer IKS is developing an AI-supported way to record symptoms over time and give conversations with a doctor a clearer, more reliable starting point.
September 3, 2026
© iStock/AleksandarNakic
An analysis by the World Economic Forum and the McKinsey Health Institute found that women live longer than men but spend roughly 25 percent more of their lives in poor health. Closing that gap would be equivalent to more than 500 additional days of good health per woman over a lifetime [1].
Part of the problem lies in the evidence base itself. Of more than 650 research papers reviewed for that analysis, close to half reported no data on sex-based differences. For ischemic heart disease and migraine, only about one in ten clinical trials reports sex-specific results [1].
The consequences show up in everyday care. Endometriosis affects roughly 10 percent of women of reproductive age, yet the delay between first symptoms and diagnosis is still measured in years, with a systematic review reporting typical delays of seven to ten years [2].
Diagnostic delay has many causes. One is structural, and in principle solvable: symptoms fluctuate across weeks and months, while an appointment captures a single moment. Everything in between — pain patterns, fatigue, sleep, cycle phase, medication — is recalled through memory, and important details are often lost.
An AI system trained on a literature that underreports women learns that imbalance along with everything else and reproduces it at scale. A 2025 study at the London School of Economics shows the risk precisely. Researchers took real long-term care records for 617 people, created gender-swapped versions, and had several language models summarize both. Same person, same needs, one attribute changed. One widely used open model consistently produced summaries in which women's health needs were described less directly and downplayed more often than men's [3].
The failure was subtle. The models applied a statistical expectation about who is likely to be seriously ill — an assumption that may appear reasonable on average but can be wrong for an individual. Errors of this kind stay hidden inside an overall accuracy score. They become visible only through subgroup evaluation and systematic bias testing, which is why that work belongs in the earliest stages of development
SympATA — Symptom Analysis and Tracking with AI — is a prototype from the Trustworthy Digital Health team at the Fraunhofer Institute of Cognitive Systems IKS. It has a deliberately focused purpose: to help bridge the gap between what a person experiences over time and what reaches the doctor during an appointment.
Users can complete simple daily check-ins and create diary entries in their own words, either by writing or speaking. This matters because predefined checklists often remove meaningful details. Checklists can also become repetitive, causing users to stop tracking. Free text preserves more detail but is harder for software to process. A compact language model, that is small enough to run locally, was therefore fine-tuned specifically for symptom extraction. That work was conducted in cooperation with Fraunhofer IIS as part of the DSgenAI project (Digital Signal Processing Using Generative Artificial Intelligence). Its task is limited: read an entry and identify the symptoms described, keeping as close as possible to the writer’s own wording.
The model learned from real diary entries contributed by women in a pilot study run by Fraunhofer IKS. Two researchers independently annotated the entries and a gold-standard dataset was produced that was then used to generate a much larger synthetic training set. In testing, the model performed close to a far larger commercial model. Because it is small, it can run on infrastructure under our control. Sensitive symptom diaries therefore do not need to be sent to an external AI service. A custom model of this kind can sit as a backend component within a more complex system for symptom capture and processing. That is what SympATA is built on: symptom tracking in free text/voice, without compromising either detail or data privacy.
A generative system in this setting has to listen to the body and mind its words. Both halves are design requirements.
Listening to the body means treating cycle phase, sleep, stress, and medication as clinically meaningful context. This is precisely the information that averaging has historically discarded, and the same loss that produced the evidence gap in the first place. SympATA is built to look for patterns across entries and to show the entries behind each pattern, such as whether a symptom appeared more often in a particular cycle phase or alongside poor sleep.
Minding its words means the summary contains only what the patient reported. Every statement traces back to a dated entry, in the patient's own terms, with no inferred symptoms and no diagnostic language dressed up as description. That constraint is not automatic; a model asked to summarize will reach for clinical shorthand unless it is trained and checked not to. A doctor can check any line against its source, and the patient can review what will be shared.
Such constraints are not only safety measures; they are essential for usefulness. A summary a doctor cannot check is a summary a doctor will not trust, and an untrusted tool changes nothing about the appointment.
For the woman using it, the gain is immediate: months of scattered experience arrive at the appointment as a clear record, and the work of remembering and re-explaining moves off the patient. For the doctor, the appointment starts from better information. Neither of those gains asks anyone to trust a machine with a decision.
The longer-term gain is larger. The evidence gap described at the start persists partly because women's day-to-day health has never been recorded systematically. Tools that capture it well, with consent and proper protection, begin to generate the longitudinal information the field has always lacked. The annotated diary corpus built for this work is a small instance of that, in a domain where datasets of the kind are close to nonexistent.
The remaining questions are the ones Fraunhofer IKS asks of every medical AI system: how it is validated and against what, whether its behavior can be explained to doctors, patients, and regulators, and where it sits under the EU AI Act and medical device regulation. In a field where confident-sounding output comes easily and reliable evidence stays scarce, that assurance work is where the real value sits.
Most of the failures described here trace back to information that existed and never arrived. That is a problem worth pointing careful technology at, so that the conversation between a woman and her doctor can start from what actually happened.
The purpose of the SympATA app is to support medical professionals, not to replace medical expertise — or the doctor. Assessment, diagnosis, and treatment decisions remain entirely with the doctor. Clinical judgment stays where it belongs.
[1] World Economic Forum and McKinsey Health Institute, Closing the Women's Health Gap: A $1 Trillion Opportunity to Improve Lives and Economies, 2024.
[2] P. De Corte, M. Klinghardt, S. von Stockum, and K. Heinemann, 'Time to diagnose endometriosis: current status, challenges and regional characteristics — a systematic literature review', BJOG: An International Journal of Obstetrics & Gynaecology, vol. 132, no. 2, pp. 118–130, 2025.
[3] S. Rickman, 'Evaluating gender bias in large language models in long-term care', BMC Medical Informatics and Decision Making, vol. 25, no. 274, 2025.
