Sofia was six years old in 2019 when her parents received the result of a genetic test that sought to explain her rare symptoms, including seizures, developmental delay, and an irregular heartbeat. The laboratory found a change in a gene, but classified it as "benign variant, with no clinical significance," leaving the family without an answer. Five years later, a doctor re-analyzed the same result, benefiting from thousands of genomes sequenced in other European hospitals and new studies published about that gene. The "benign variant" of 2019 was reclassified as the cause of Sofia's disease, although her DNA had not changed, only the scientific knowledge about it.
The article compares this situation with the book "Sophie's World" by Jostein Gaarder, where the protagonist gradually discovers that each new idea forces her to revise what seemed certain. In the life and health sciences, the meaning of biological information changes continuously with the advancement of knowledge, unlike a postal code that always means the same thing. An AI system trained with 2019 data would continue to classify that gene as harmless, emphasizing that the data used to train AI models in healthcare needs permanent curation by specialists, not a job done once and forgotten.
Rare diseases present a particular problem for the development of AI models: they affect few people, but each patient generates an enormous amount of information. A hospital may have only three to twenty similar cases, but AI models need thousands of examples to distinguish real patterns from mere coincidences. The obvious solution would be to gather data from all European hospitals in a single center, but health data is highly sensitive and protected by law, creating privacy risks if taken out of the country of origin.
This is where federated data sharing comes in: instead of sending complete patient records to a central vault, each hospital keeps the data locally and only answers specific questions. An AI model "travels" between different secure data centers, learns in each one, and returns without the original data leaving the premises. However, federation alone does not guarantee privacy, as poorly designed models can memorize unusual cases. Additional layers of protection are needed, including controlled statistical noise, rules that prevent results based on few patients, and secure research environments (Trusted Research Environments) that function as monitored reading rooms. INESC-ID, through its thematic line Life and Health Technologies, has been working on building these secure environments in partnership with national hospitals and universities, essential for Portugal to develop truly personalized medicine.
