The study involved 30 research groups across six countries — including Australia, the United States, Canada, Spain, Greece, and Austria — and analysed 7,525 cancer sample
Instead of moving patient data, each participating institution trained an AI model using their own dataset. Only the model insights and learnings — not the data itself — were then shared with a central server, which compiled the findings into a single, highly accurate diagnostic tool
In a landmark advancement for cancer research and treatment, an international team of scientists has developed a novel artificial intelligence (AI) technique that enables the secure, large-scale analysis of cancer samples while protecting patient privacy.
The breakthrough, led by researchers at Australia’s Children’s Medical Research Institute (CMRI), uses a federated deep learning approach to analyse proteomic data — protein profiles that provide a biological fingerprint of cancer — from over 7,500 samples collected worldwide.
The findings, published in the journal Cancer Discovery, mark a significant step forward in the pursuit of personalised cancer therapies, allowing doctors to better match treatments to the individual characteristics of each patient’s tumour.
“This is a game-changer for global cancer collaboration,” said Professor Roger Reddel, CMRI Director and Head of the Cancer Research Unit. “It was a very exciting moment when we first saw that results from data with highly restricted access were just as accurate as results obtained when the data was stored in one place.”
The study involved 30 research groups across six countries — including Australia, the United States, Canada, Spain, Greece, and Austria — and analysed 7,525 cancer samples. These samples were used to study proteomes, or protein profiles, which give scientists crucial insights into how different cancers behave and respond to treatments.
However, international collaboration in this field has long been hampered by two major obstacles: Strict patient privacy laws, which prevent the sharing of sensitive health data across borders.
Technical differences in how laboratories process and analyse samples, making it difficult to integrate data from different sources.
To address these challenges, the CMRI team turned to federated learning, an emerging branch of AI that allows models to be trained on local datasets without transferring the raw data.
Instead of moving patient data, each participating institution trained an AI model using their own dataset. Only the model insights and learnings — not the data itself — were then shared with a central server, which compiled the findings into a single, highly accurate diagnostic tool.
This decentralised approach not only safeguarded patient privacy but also enabled the integration of data obtained through different lab techniques, further improving the AI system’s diagnostic performance.
“Our goal with the ProCan research programme is to build proteomic tests that help clinicians choose the most effective treatments for their cancer patients,” Reddel explained. “By overcoming these long-standing data barriers, we’re now significantly closer to making that vision a reality.”
The potential implications are wide-ranging. As cancer becomes increasingly understood as a diverse group of diseases rather than a single condition, personalised medicine is considered key to improving survival rates and treatment outcomes.
Using proteomic data — the proteins produced by cancer cells — researchers can classify tumours more precisely and predict which therapies are most likely to be effective for each patient.
This study not only advances the science behind personalised treatment but also sets a new precedent for secure, large-scale AI collaborations in global healthcare. “This AI-driven approach opens the door to unprecedented levels of cooperation without compromising patient confidentiality,” Reddel added. “It’s a win for science, medicine, and ethics.”
