Epileptic seizure or not? New deep learning model has the answer
Technology to detect epileptic seizures may soon be more efficient and reliable than ever before, thanks to a new machine learning model developed by researchers at the University of British Columbia.
While some seizures take the form of violent convulsions or a loss of consciousness, others produce less visible symptoms, such as staring spells or lapses in awareness. This makes them more difficult to detect or diagnose at an early stage, increasing patients’ risk of harm.
“Having epilepsy doubles or triples your chances of premature death, so it’s critical for patients to get help as soon as a seizure begins,” says Ramy Hussein, a doctoral student in electrical and computer engineering and the lead author of the study. “Our method can detect seizures more dependably than current state-of-the-art approaches, even under noisy, real-life conditions.”
Epileptic seizures are bursts of electrical activity in the brain that are typically diagnosed by identifying specific patterns in electroencephalogram (EEG) signals — the spikes and waves that represent voltage fluctuations within the brain. EEG-based epileptic seizure detection devices include personal EEG headsets, which are often worn 24 hours a day by patients with medication-resistant epilepsy, and machines used in brain research institutions and hospitals.
But visually identifying epileptic seizure-related patterns in EEG signals is a labour-intensive, time-consuming process, and most automatic detection methods are hampered by the variability and noisiness of EEG data. Seizure patterns change not only from person to person, but also over time for the same person. What’s more, they can be corrupted by everything from muscle movements and eye blinks to nearby power lines and other sources of environmental noise.
So the UBC team, led by electrical and computer engineering professors Rabab Ward and Jane Wang, developed an automatic detection method that effectively filters out the noise and otherwise accounts for the variability in EEG signals, potentially making it suitable for the clinical diagnosis of epileptic seizures. By extracting the most robust and distinctive EEG attributes associated with seizures, it is able to detect seizures with 90 to almost 100 per cent sensitivity, specificity and classification accuracy under real-life conditions — that is, in the presence of common signal contaminants like muscle activity, eye movement and ambient noise.
In the absence of such contaminants, the method performed perfectly — with 100 per cent sensitivity, specificity and accuracy. The UBC investigators conducted their experiments using EEG signals from a popular public EEG database created by Bonn University, which were then distorted using different combinations and intensities of noise for the purposes of the real-life simulations.
The team is now working to improve the performance of their model by training it on a larger EEG dataset and modifying it to accommodate multi-channel EEG systems. They are also incorporating long-term EEG signals into their experiments in order to identify key pre-seizure EEG signals — information that could potentially help predict future attacks.
“Detecting epileptic seizures is clearly very important, but being notified about a seizure before it actually occurs may be even more helpful to patients,” says Ward. “It would enable them to prevent or minimize danger by taking the necessary medication, moving away from sharp or hard objects, or otherwise arranging for timely help from a caregiver, who could ensure that their airway remains clear during the seizure or obtain medical help if any complications arise.”
Epilepsy is one of the most common neurological disorders in the world, affecting over 70 million people globally — 30 to 40 per cent of whom do not respond to medication. The leading cause of death among people with epilepsy, Sudden Unexpected Death in Epilepsy, is believed to relate to abnormal breathing or heart rhythms and usually occurs during or after a seizure.
The study, which appeared in Clinical Neurophysiology, was conducted by Hussein, Ward, Wang and Hamid Palangi, currently a researcher at Microsoft Research AI.
Feature photo credit: "Insight picture" by Chrissshe is licensed under CC BY-SA 4.0