Reading Before Designing: What Our Investigate Phase Taught Us
AIRIS Team · 20–10–2026
Evidence first, then design. Exploring the evidence, understanding the journey.
Three weeks ago we wrote that framing a problem well comes before building anything. The Engage phase gave us a Big Idea, an essential question and a challenge. The Investigate phase is where we tried to prove ourselves wrong: we asked 91 guiding questions, read 66 recent articles, compared seven existing tools, mapped public datasets for each axis, and ended with an objectives tree for gAIa.
No model exists yet. What exists is a clearer picture of the problem, and a list of things we now refuse to promise.
Where we start: an unequal journey
About 1 in 6 people worldwide experience infertility (WHO). In Tunisia, the figure we found for couples is around 15%, and we treat it with caution because the available figures are not always complete. What matters more than the number is how the journey feels: limited access to care, social stigma, and responsibility that is rarely shared equally between partners.
How we asked: 91 questions, 11 families
We started from our essential question and split it into eleven families, from the problem itself to the people around it:
- Problem and context: what fertility and infertility mean medically, how a couple’s work-up really unfolds, and who is involved.
- Causes and consequences: medical causes in men and women, why infertility stays unexplained, what delays couples, and what it does to their emotional health and relationship.
- State of the art: which tools exist, who they are built for, how they were validated, and what we can learn from solutions that disappointed.
- Ethics and law: privacy, consent, bias, uncertainty, and the point at which a tool becomes a medical device.
- Data, tools, resources: which datasets exist, which methods fit each data type, what it costs, and which experts can review our work.
- Environment and stakeholders: exposures such as heat or pesticides, and what couples, clinicians, laboratories and associations each need.
Three questions shaped everything after them:
- Is the man assessed as early and as thoroughly as the woman, and why not?
- What makes a medical report hard for a couple to understand?
- How can a language model write a report that uses only measured values, cites its sources and never invents a number?
What we did and what we read
Our inquiry moved in four steps: discovering the challenge, formulating guiding questions, researching clinical and technical evidence, then analysing what we found. Along the way we put names on the vocabulary couples are given: subfertile, fertile, infertile and sterile are not synonyms, and the usual chance of a natural pregnancy is roughly 20 to 25% per cycle in a fertile couple. We also mapped the assisted-reproduction ladder (ovulation induction, IUI, IVF, ICSI), because a report that does not say where a couple stands on that ladder is hard to act on.
Then we read. Between 2024 and 2026 we went through 66 articles, and 47 of them (71%) are high-level evidence: 15 meta-analyses, 14 systematic reviews and 18 randomised trials. The other 19 are cross-sectional studies and case reports, which we use for context rather than proof. Coverage is balanced on purpose: 12 articles on women, 11 on men, and 6 on Tunisia.
Six studies shaped our objectives most:
The Tunisian data stayed with us. In a study of 60 couples (120 people) at Aziza Othmana Hospital in Tunis, anxiety symptoms reached 45%, depressive symptoms 25% and sexual dysfunction 39%. Behind the figures are three experiences that came up again and again in our sources: guilt and blame, silence and taboo, and doubt and confusion.
What already exists, and what data we can use
We compared seven tools (Tanit AI, Leeaf, iYoni, Flo, Embie, ExSeed, BirthAI) on eight features: AI assistant, cycle and symptoms, result insights, ART/IVF support, male fertility, female conditions, image or video analysis, and connection with professionals. In our grid, each tool covers part of the journey well. Tanit AI is strong on assisted reproduction and clinic support, Flo on cycles, ExSeed on semen analysis. None of the seven fully covers female conditions, and none brings the man’s and the woman’s examinations into one explained picture. That is the space we want to work in, and we say it as an observation from our grid, not as a claim that nothing else exists.
Get Ferdaouskachouri’s stories in your inbox
Join Medium for free to get updates from this writer.
We also took one lesson from the trial above: a tool that is faster but not shown to be better has to say so plainly.
On data, we catalogued public datasets for each axis:
Two honest limits came out of this step. First, these datasets do not describe the same patients, so we will link our modules by clinical concepts and never pretend they belong to one person. Second, some key data are missing publicly, for example varicocele images, or ultrasound and hormones for the same women.
For methods, the pattern is simple: YOLOv8 with ByteTrack for sperm and embryo videos, EfficientNet for microscopy, ultrasound and MRI foundation models, GRU-D or a Temporal Fusion Transformer for cycles, XGBoost for tables, and a local language model for the report. Everything is paired with explainability (Grad-CAM, SHAP), anonymisation and access control.
What the investigation revealed
We merged four layers of evidence into one synthesis: the scientific literature (global cohorts and peer-reviewed AI capabilities), clinical guidelines (WHO standards, EAU protocols), diagnostic methods (the limits of visual exams, the promise of multimodal data) and patient perspectives (stigma, emotional toll, the need for localisation). Five findings came out of it:
- Results are separated and hard to read. Men’s and women’s examinations are done apart, depend partly on the examiner’s judgment, and are rarely explained to the couple.
- The burden is not shared equally. Blame, silence and confusion weigh differently on women and men, and Tunisian couples report high anxiety.
- AI is promising but young. Studies on sperm morphology and motility are small and weakly validated externally, and embryo-selection AI was faster but not shown to be as good.
- The data are scattered. Public datasets exist for each axis, but they come from different people, and some data are missing.
- Language matters. An explanation only helps if the couple can follow it, in their own language.
From insights to an objectives tree
These findings became the objectives of gAIa: one overall goal, three branches, and a shared report on top.
Each module works alone, because a couple may come with only one partner’s results.
What we will not promise
Investigating also meant writing our limits down before anyone else does:
- gAIa supports a decision; it never diagnoses, never says that X causes Y (only that X is associated with Y), never predicts anything about a child and never recommends a treatment.
- Every generated image or curve is labelled SYNTHETIC, so it cannot be mistaken for a measurement.
- Some datasets are small (VISEM covers 85 men, mcPHASES 42 people), and most data come from Europe, North America and Asia, not Tunisia.
- Our modules do not share the same patients, and nothing has been clinically validated.
- No patient data will be used without an ethics committee, consent and anonymisation.
gAIa is a university research prototype, not a medical device.
What comes next
The Act phase starts with the first baseline models for each axis, then the integration into one report. In parallel we will ask clinicians, laboratories and associations to review our medical content, run our first interviews and an anonymous survey, and apply for access to the datasets that require an agreement.
If you work in reproductive health, or you have lived this journey, we would like to hear what you find hardest to understand in your results. Follow our progress on Instagram at Airis.labs.
We are building gAIa to help you understand yourself, understand your body, and understand each other.
