Image
UVA Researchers Develop AI Tool to Detect Seizures More Accurately
Researchers at the University of Virginia developed an AI system that can identify the brain-wave patterns of absence seizures from EEG recordings far more accurately than previous automated methods, potentially saving researchers hundreds of hours of manual analysis.
Unlike the convulsive seizures many people associate with epilepsy, absence epilepsy is marked by brief lapses in consciousness. A person, often a child, may stare blankly for a few seconds before resuming normal activity, sometimes without realizing a seizure has occurred. These episodes can happen dozens or even hundreds of times a day, disrupting learning, attention, and everyday life.
Since these seizures usually have few visible signs, doctors use EEGs, which record the brain’s electrical activity, to spot the spike-wave patterns that confirm an absence seizure.
Today’s EEG systems can record brain activity nonstop for days or weeks, creating huge amounts of data that researchers must review by hand. Finding the specific brain-wave patterns linked to absence seizures, called spike-wave discharges, often means experts must look through hundreds of hours of recordings. This makes the process slow, tiring, and hard to expand.
Researchers at the University of Virginia have developed an artificial intelligence model that could dramatically reduce that burden. In a new paper published in Scientific Reports, a collaborative team from the UVA School of Data Science and the School of Medicine describes an AI approach that automatically identifies the precise timing of absence seizures in EEG recordings with greater accuracy than existing automated methods.
The team trained their model on almost 1,000 hours of EEG recordings, all carefully labeled by hand. This created one of the largest datasets of its kind. They tested 16 different machine learning methods before building a model that worked better across different people and recording situations.
“We set out to find which existing approaches transfer to dense temporal segmentation of single- channel EEG, and then adapt to the one that worked," said Saurav Sengupta, a data scientist who led the project. "The real test was whether it could detect absence seizures in mice it had never seen before, as the model was trained on eight animals and evaluated on ten completely separate ones, because generalizing across individuals is exactly where earlier methods could break down. The result is a foundation where we can build more capable models rather than an end point."
This new model is different from earlier ones because it was built to recognize the natural differences in EEG recordings from different people. By training the AI with a wide range of real brain signal variations, the researchers made it better at telling true seizures apart from background activity and noise. It is 29% more accurate than the best automated method used previously.
The impact of this research goes beyond just this study. Automated tools like this could help researchers handle much larger datasets, speed up studies of epilepsy and other brain disorders, and cut down on time spent doing repetitive manual work.
To help others make more progress, the researchers have shared both their trained model and their labeled dataset with the public. This lets other scientists build on their work and use the approach for other brain research.
This study demonstrates the usefulness of AI in helping researchers analyze complex biomedical data more efficiently and consistently.
"Combining residual U-Net and data augmentation for dense temporal segmentation of spike wave discharges in single-channel EEG" was published on July 21, 2026, by Scientific Reports and co-authored by Saurav Sengupta, Suchetha Sharma, and Don Brown from the UVA School of Data Science and Scott Kilianski, Sakina Lashkeri, Ashley McHugh, and Mark Beenhakker from the UVA School of Medicine.
