A University of Technology Sydney AI model posted the highest numerical scores for separating ischemic stroke cases from other participants in UK and U.S. datasets, while its clinical benefit remains untested.
The study, published online Sept. 8, measured that separation using area under the receiver operating characteristic curve (AUROC). The model scored 0.8109, 0.7843 and 0.7468 across three UK Biobank comparison groups, and 0.7251 in the U.S. All of Us cohort. Each was the highest numerical AUROC among the models evaluated in that setting. The scores measure how well the model distinguished groups in the data; they do not show that using it prevents strokes.
The researchers built the model with a transformer, an AI architecture also used in large language models. It takes account of where genetic variants occur in the genome, an approach the authors say may capture relationships that conventional additive genetic risk scores miss. The model’s public documentation lists 477 genetic markers and family history of ischemic stroke as inputs. The study also identified 18 genomic regions associated with the condition.
In a separate analysis of new strokes, participants in the highest score group had a hazard ratio of 1.21 compared with those in the lowest group. That figure describes a relative difference in the rate of strokes over time, not any one person’s absolute chance of having one.
Ischemic stroke occurs when a blockage disrupts blood flow in the brain. The World Health Organization estimates that stroke of all types caused 6.8 million deaths worldwide in 2023, making it the second leading cause of death that year. “Genetics are only one part of stroke risk,” said UTS Distinguished Professor Jie Lu, who directs the Australian Artificial Intelligence Institute.
The University of Technology Sydney says the work could give patients and families clearer information about susceptibility. Its team is developing a broader assessment that would consider clinical and lifestyle information alongside genetic data. The current findings do not establish whether such information would change care or prevent a stroke.
The study’s authors say further research should follow people over time and include populations with more varied ancestries. Dr. Kairui Guo, the study’s corresponding author, said a preventive clinical trial has been proposed.
