Deep neural networks applied to interictal EEG to improve the diagnosis of epilepsy
Epilepsy is a debilitating disease affecting 1% of the world’s population. It is characterized by an enduring propensity to recurring seizures. One of the cornerstones of the diagnosis and management of epilepsy is the electroencephalogram (EEG). The EEG is a recording of the brain’s electrical activity, typically for 30-minutes, and is interpreted by a neurologist in search of interictal spikes. These “spiky”, asymptomatic discharges are a strong predictor of seizure risk and occur between seizures; unfortunately, they occur infrequently, and a routine EEG will manage to capture interictal spikes in less than half of patients with epilepsy. In the absence of interictal spikes, neurologists cannot distinguish between the EEG of patients with and without epilepsy. In the past decades, researchers have turned to computers to identify subtle patterns in the EEG of patients with epilepsy, patterns that are hidden to the naked eye. Deep learning (DL) is a type of artificial intelligence that is particularly performant in identifying subtle patterns in complex data. Could DL effectively identify epilepsy on the EEG without interictal spikes?
This project aims to: 1) construct a large database of anonymized EEG recordings, reviewing the patient’s medical files to verify if the patient is epileptic or not, and 2) implement and train a DL algorithm capable of recognizing the brain signal of patients with epilepsy, even in the absence of interictal spikes.
This project could have a significant impact on clinical practice. The identification of new biomarkers of epilepsy could increase the capacity of the EEG to detect this disease and allow to identify and treat these patients earlier. Moreover, throughout the project, emphasis will be put on sharing data and computer code to other scientists to ensure our analyses are reproducible and benefit other research groups.
Elie Bou Assi , Université de Montréal
Partners and Donors