How Deep Learning Techniques Can Predict Epilepsy Seizures

Australian researchers have developed a proof-of-concept of a wearable device that can predict the onset of seizures in patients with epilepsy.


Australian researchers have developed a proof-of-concept of a wearable device that can predict the onset of seizures in patients with epilepsy.

The retrospective study of data gathered from 10 patients simulated real-life scenarios by deploying them on a brain-inspired chip, according to  Stefan Harrer, PhD, manager of brain-inspired computing research for IBM Research Australia. As part of that simulation, the researchers used deep learning, creating algorithms that would learn which brain signals unique to an individual would preempt seizures.

“People with uncontrolled epilepsy often live with uncertainty about when a seizure is going to occur,” they wrote in the study. “It is this uncertainty that can lead to difficulties with daily life activities, such as driving, working, or even socializing, and may expose an individual to unnecessary danger. By presenting information about when a seizure is likely to happen, we hope to restore a degree of perceived control.”

Epilepsy is the fourth most common neurological disease and affects people of all ages, according to the Epilepsy Foundation. With 150,000 new cases every year, 3.4 million people in the United States live with active epilepsy.

The brain chip yielded a 41.3% mean improvement in seizure prediction over chance, which is 9% less than the mean results obtained on a high-performance computer. It also warned of an upcoming seizure during 31.7% of the time that the system was in operation.

Previous studies failed to predict seizures for all patients in this dataset because they used conventional machine learning, according to Harrer. In those studies, an expert clinician would interpret the EEG data and manually identify brain signals that could be used to train the algorithm on an individual's brain states leading up to a seizure.

“This new research uses deep learning which self-learns automatically and was able to uncover signs that were invisible to the human eye — meaning some level of prediction was possible for each patient,” Harrer wrote in an email to Qmed. “It demonstrated the ability to train systems to individual patients.”

Deploying the system on a brain-inspired chip, the new study showed how the system could work as a wearable, Harrer added.

"This is an important breakthrough as previous studies were achieved on high-power computers,” he wrote.

The study’s results provide a proof-of-concept of a robust, real-time, low-power seizure prediction system that offers patients the ability to personally configure the seizure warning system, according to the researchers. Predicting every seizure is a ways off, however.

“There is still work to be done in this research to improve prediction across all patients and reduce false positives before it could be applied to real patients,” Harrer wrote. “Further - To ensure a system could be scalable to more patients, algorithms would need to be able to be trained on data from outside the skull, rather than inside. We cannot confidently put a timeline on this.”

Beyond real-time seizure prediction, the system could also be applied to a closed-loop therapeutic device for titrating therapies such as neuromodulation or acute drug delivery, Harrer added. This could reduce the number of drugs or therapeutic stimulation delivered to the patient, reducing the treatment burden and side effects, he wrote.

500 characters remaining