Scientists have developed an artificial intelligence model that reads a person's DNA to flag genetic predispositions to stroke, the second leading cause of death worldwide. The tool, built on the same transformer architecture that powers large language models like ChatGPT, could transform how clinicians assess risk and tailor prevention strategies.
The research, led by a team at the University of Technology Sydney (UTS), focused on ischaemic stroke, the most common form, caused by a blood clot blocking blood flow to the brain. By training the model on genetic variations, the team identified 18 variants linked to the condition, with key genes such as EDEM2 (active in the brain cortex) and KANK2 (found in whole blood) emerging as major contributors.
“When someone has a stroke, particularly at a young age, they often want to know whether other family members are also at risk,” said Kairui Guo, the study's corresponding author. “These new AI tools will provide the ability to better help families understand that risk.”
Genetics account for roughly 38% of an individual's stroke risk, according to earlier research. Yet traditional risk assessments have largely overlooked this factor. “Traditionally, in thinking about stroke risk, genetics hasn't been a main consideration; however, this study highlights the significant impact of genetic factors,” Guo added. He noted that processing the sheer volume of genetic data needed for a personalised assessment would be impossible without AI.
Beyond traditional risk scores
Until now, scientists have relied on polygenic risk scores (PRS), a statistical method that estimates disease risk based on the combined effect of many genetic variations. However, PRS fails to account for how different variants interact with one another or how their specific positions along the DNA sequence influence overall risk. The new transformer-based model overcomes these limitations by learning complex patterns from the data.
The team tested the model on two major medical datasets, together comprising more than 40,000 individuals. It successfully categorised people into low, medium, and high genetic risk groups. Over time, those in the high-risk group were 21% more likely to suffer a first-time ischaemic stroke compared with those in the low-risk group.
“AI can be used to find associations between human genes and disease, and build personalised risk assessments, to help prevent a potentially life-changing or fatal event,” said Jie Lu, director of the Australian Artificial Intelligence Institute.
Lu outlined the next step: a multimodal AI tool that combines genetic risk with other data, such as health and lifestyle surveys, scans, and imaging. “Genetics are only one part of stroke risk. The ability to combine a genetic score with clinical and lifestyle information, including established risk factors such as blood pressure, age, atrial fibrillation, type 2 diabetes and cholesterol, could enhance monitoring and prevention,” she said.
The approach echoes other recent innovations in stroke care across Europe, such as graphene sensors that map stroke damage in real time, developed by a Spanish-led team. Meanwhile, researchers are also exploring unconventional rehabilitation methods, including birdwatching as a therapy for stroke-related aphasia.
While the UTS model is still in the research phase, its potential for personalised medicine is significant. By identifying individuals at high genetic risk early, clinicians could recommend targeted lifestyle changes or preventive treatments before a stroke occurs. The team hopes to refine the tool and eventually integrate it into routine clinical practice, offering a new layer of protection against a disease that claims millions of lives each year.
