Researchers have developed an artificial intelligence model that identifies 18 genetic variants linked to ischaemic stroke risk, offering a new path toward personalized health assessments. People with the highest genetic risk scores face a 22 per cent higher chance of stroke compared to those in lower-risk groups.
University of Technology Sydney Researchers Train Transformer Model
Scientists at the University of Technology Sydney (UTS) built the diagnostic tool using transformer architecture, the same underlying technology that powers large language models like ChatGPT. The system processes thousands of genetic variations simultaneously, evaluating their exact chromosomal locations and potential biological interactions. Yue (Eric) Yang, who recently completed a PhD in artificial intelligence at UTS and served as the study’s first author, explained the system’s performance against existing benchmarks. Yang stated that the model proved more accurate at distinguishing individuals who experienced a stroke from those who had not when tested on established health datasets.
The research team evaluated the algorithm using genetic and medical records from the UK Biobank alongside health data from the large, US-based All-of-us database. Beyond overall risk stratification, the analysis highlighted the EDEM2 gene within brain cortex tissue as a key contributor to stroke susceptibility. This discovery points toward specific biological mechanisms that future medical research can target.
University College London Virtual Trials Simulation
Complementing the genetic findings, researchers from University College London (UCL) developed an AI-powered virtual trial platform designed to refine treatment recommendations for ischaemic stroke. Published in Nature, the UCL study involved 100 million biologically realistic simulations to demonstrate how traditional randomized controlled trials may miss vital individual differences in brain biology. Dr Dominic Giles, lead author of the UCL study from the UCL Queen Square Institute of Neurology, noted that diseases affect individuals uniquely, yet clinical decisions frequently rely on crude metrics such as symptom duration or overall lesion size.
The UCL virtual trial framework combined large-scale maps of brain connectivity, genetic expression data, receptor distribution figures, and more than 4,000 real acute stroke lesion maps aligned to a common brain template. This architecture enabled researchers to simulate realistic disability patterns and biological treatment responsiveness. Although the complex models incorporated noisy data, they delivered more accurate personalized treatment predictions than the simplified models traditionally utilized in clinical trials.
Combining Genetics With Clinical Risk Factors
Jie Lu, director of the Australian Artificial Intelligence Institute and an Australian laureate fellow in the UTS Faculty of Engineering and IT, emphasized that DNA represents only one component of a patient’s overall susceptibility. Lu explained that combining genetic scores with established clinical and lifestyle metrics—such as blood pressure, age, atrial fibrillation, type 2 diabetes, and cholesterol—enhances monitoring and prevention protocols.
Building on this strategy, the UTS team is currently developing a multimodal AI tool. This upcoming system merges genetic risk scores with health surveys, lifestyle data, medical scans, and imaging to evaluate susceptibility not only to stroke but also to conditions such as endometriosis.
Frequently Asked Questions About AI Stroke Risk Models
How much higher is the stroke risk for people in the top genetic category?
Individuals with the highest genetic risk scores identified by the UTS model have a 22 per cent higher chance of stroke than those in the lower-risk group.
Which specific gene was identified as a key contributor to stroke susceptibility?
The analysis pinpointed the EDEM2 gene in brain cortex tissue as a key contributor to susceptibility, providing a target for future biological research.
What databases were used to train and evaluate the UTS artificial intelligence model?
The researchers evaluated the algorithm using genetic and health data from the UK Biobank and health records from the US-based All-of-us database.
How many simulations did the University College London virtual trial platform analyze?
The UCL research involved 100 million biologically realistic simulations to test personalized treatment predictions against traditional randomized controlled trial models.
Related reading