Epilepsyecosystem.org: Crowd-Sourcing Reproducible Seizure Prediction with Long-Term Human Intracranial EEG

Kuhlman L, et al. Brain, 9(1): 2619-2630.

Unpredictability of seizures is a top issue in the epilepsy community. There is both the fear of not knowing when a seizure will start and not knowing what triggers the seizure onset.

In 2013, Neurovista completed the world's first in-human clinical trial of an implantable seizure prediction device. For this study, 15 people living with uncontrolled seizures had electrodes implanted in their brain and shared their data for up to 3 years. The electrodes recorded EEG (electroencephalography) signals for 24 hours a day, 7 days a week, over a long period of time.

This implantable seizure prediction system worked well for some individuals. But for three participants, the system did not work at all.

Purpose

• Researchers wanted to look at those 3 individuals and see if better algorithms could be developed for a personalized seizure advisory system. More about this study.
• They developed a crowd-sourcing platform that shared data on the 3 individuals who had low seizure prediction performance to gather submissions on possible solutions.

Description of Study

• The goal was to see if algorithms developed by the community could improve upon those already developed from the Neurovista trial data.
• Participants could use any programming language, data processing method, or machine learning method to develop this algorithm.
• Once an algorithm was submitted, it was tested on an unreleased data set from the 3 individuals to see if the seizure-prediction algorithm was accurate.
• The researchers evaluated the following about the algorithms:
  • Their ability to predict when someone had a high or low chance of having a seizure.
  • How specific was the proportion of warning time the algorithm would give the individual. For example, if you are always in a high-warning state, you would predict all the seizures, but this would not be useful to the individual.

Summary of Study Findings

• 646 individuals in 478 teams from around the world entered the competition and submitted over 10,000 different algorithm entries over the course of 2 months.
• Different contest algorithms performed better depending on the person with epilepsy it was being tested on. This means there won't be a one-size fits all approach to seizure prediction and personalizing algorithms matter.
• For 2 out of the 3 individuals selected for the study, the new algorithms were significantly better at predicting when they were high-risk for having a seizures than the previous algorithms.
• All the winning algorithms could indicate when someone was not likely to have a seizure.
• Because of the success of this competition, the researchers have developed a platform that is exclusively dedicated to improving seizure prediction algorithms known as www.epilepsyecosystem.org. Every year, they will re-evaluate the algorithms being submitted and also test the algorithms on the full Neurovista clinical data sets.

What does this mean?

 

• There is value to inviting different people to crowd-source ideas on how to solve the problem of seizure prediction.
• In the future, seizure prediction algorithms will need to be tailored to the individual. The solution might be to test a lot of different prediction algorithms and chose the one best for each specific person.
• There may be other components in addition to EEG that could be helpful in improving seizure prediction algorithms. For example, many people living with epilepsy have specific triggers that increase their risk of seizures. Can we measure those changes in the body in addition to electrical signals (such as stress, sleep quality, etc.) to improve seizure prediction algorithms?
• We are now able to collect, process, and analyze more data than ever before. It should be possible to start merging multiple types of information, such as electrical recordings and changes in the body.

My Seizure Gauge

The Foundation has launched the My Seizure Gauge challenge to drive the seizure forecasting field forward. We want to create a minimally invasive individualized seizure gauge that would allow a person with epilepsy to monitor the likelihood of a seizure on a daily basis. Our purpose is to identify and better understand the changes in the body (in addition to EEG signals) that may occur before a seizure, within hours or days before the clinical (observable) seizure. Data collected from this initiative will also be put on the epilepsy ecosystem platform.

To learn more about the My Seizure Gauge Challenge, please contact Sonya Dumanis at sdumanis@efa.org.

Article published in Brain, September 2018.