Georgis-Yap and Khan
Zakary Georgis-Yap, (L), first author of the study, is a previous Master's student in the lab of Dr. Shehroz Khan, a scientist at UHN’s KITE Research Institute and senior author of the study. (Photo: UHN Research Communications)

By UHN Research Communications

Scientists at UHN's KITE Research Institute offer new hope for epilepsy research as they develop deep learning models to predict epileptic seizures.

Epilepsy, one of the world's most prevalent neurological disorders, affects more than 50 million people worldwide. Characterized by the sudden onset of seizures, epilepsy can lead to serious physical injury and even death.

The ability to predict the onset of epileptic seizures can significantly reduce injury and improve quality of life.

A team led by Dr. Shehroz Khan, a scientist at KITE and senior author of the study, focused on leveraging deep learning models to analyze electroencephalogram (EEG) data. EEG is a test that uses small electrodes to measure brain activity and serves as a vital tool for understanding seizure onset.

"Deep learning models are advanced computer algorithms that learn to recognize patterns and make predictions by processing large amounts of complex data," says Dr. Khan, who is also an assistant professor at the University of Toronto's In​stitute of Biomedical Engineering.

"By using these models to distinguish pre-seizure EEG patterns, we can help epilepsy patients and their caregivers anticipate seizures and take preventive measures."

Electroencephalography (EEG) can be used to measure regular brain activity as well as brain activity before, during and after a seizure. This data can be leveraged to predict an incoming seizure; however, there is considerable inter- and intra-patient variability in pre-seizure brain activity, which makes it challenging to develop seizure prediction approaches. (Photo: Getty Images)

Using a combination of supervised and unsupervised deep learning approaches, the researchers trained the learning models to identify subtle changes in brain activity preceding seizures.

"Supervised deep learning involves using labelled data where seizure occurrence is known," explains Zakary Georgis-Yap, a previous Master's student in Dr. Khan's lab and first author of the study. “On the other hand, unsupervised deep learning allows the model to learn predictive patterns from unlabelled data on its own.

"The advantage of unsupervised learning models is that they do not require comprehensively labelled data – which can be challenging and time-consuming to obtain."

To evaluate the effectiveness of their models, the researchers conducted extensive testing on two large s​eizure datasets containing EEG-recorded data from 40 patients.

The results of the study were promising, showcasing the feasibility of both supervised and unsupervised approaches in seizure prediction.

However, prediction results for both models varied across datasets, patients and learning approaches, highlighting the considerable variability in pre-seizure brain activity between individuals.

"While there is still work to be done, our research represents a significant step forward in the field of epilepsy management," concludes Dr. Khan. "By harnessing the potential of deep learning, we have the opportunity to develop personalized therapeutic interventions and ultimately save lives."

This work was supported by the Natural Sciences and Engineering Research Council of Canada, the Data Science Institute at the University of Toronto and UHN Foundation.

Zakary Georgis-Yap and Dr. Shehroz Khan contributed equally to the study.

Back to Top