NIH. US: For millions of people with epilepsy, life comes with too many
restrictions. If they just had a reliable way to predict when their next
seizure will come, they could have a chance at leading more independent
and productive lives. That’s why it is so encouraging to hear that researchers have
developed a new algorithm that can predict the onset of a seizure
correctly 82 percent of the time.
Until recently, the best algorithm was
hardly better than flipping a coin, leading some to speculate that
seizures are random neurological events that can’t be predicted at all.
But the latest leap forward shows that seizures certainly can be
predicted, and our research efforts are headed in the right direction to
make them even more predictable. The other big news is how this new
algorithm was developed: it’s the product of a crowdsourcing
competition.
Crowdsourcing builds on the recognition
among software developers in the mid-1980s that a crowd of users, not
just the guy writing code in a cubicle, often knows best how to design
existing products or work out the bugs in existing ones. As the
credibility of the crowd has grown in recent years, an initial wave of
biological crowdsourcing competitions has appeared online. The
competitions often pit mathematicians, computer scientists, and other
capable big-data crunchers against each other or organized into teams.
Their challenge is to solve a problem to which many often arrive at
their computer screens short on expertise but long on innovative ideas
to cut through the complexity. For organizers, the key is to model the
right problems that lend themselves to crowdsourcing, attract the right
teams, and offer the right incentives for them to drill down to an
answer.
That was the idea behind the American Epilepsy Society Seizure
Detection Challenge. The challenge was launched last August on the web
site Kaggle.com,
a well-known online platform for data prediction competitions, and
co-sponsored by the American Epilepsy Society, NIH’s National Institute
of Neurological Disorders and Stroke (NINDS), and the Epilepsy
Foundation. The challenge involved two distinct contests: detection and
prediction of seizures. A total of 504 teams from all over the world
participated in both contests, which remained open until early November.
The teams analyzed a huge data set detailing the electrical activity
in the brains of people while under evaluation for surgery to treat
their epilepsy. They also had an even larger data set from studies with
dogs, whose epilepsy closely resembles that seen in people.
All were from previous work involving collaborators from the
University of Pennsylvania, the Mayo Clinic, and the University of
Minnesota. It was their idea to hand off the data for the crowdsourcing
competition.
The goal was to identify premonitory signatures of electrical
activity in the brain during the hour prior to a seizure. In the
detection contest, the team that identified the earliest changes in
brain activity that led to a seizure with the fewest false alarms took
home the prize. In the prediction contest, the team that generated a
predictive signature of changes in brain activity that led to a seizure
with the fewest false alarms won.
After a few months, the winning prize in the detection contest went
to a computer engineer from Australia named Michael Hills. He competes
in online contests to test his skills on the side in machine learning
and digital signal processing. Hills used a special algorithm to
classify various aspects of localized electrical field potential in the
brain.
The seizure prediction contest was even more challenging, and it came
down to a tight, seven-team race. But two of the teams—one from
Australia, and the other from the United States—merged their talents
down the stretch to take the prize by predicting 82 percent of seizures.
Interestingly, neither team ever has met face to face.
Just for the record books, the US team members are Drew Abbot, a
software engineer, and Phillip Adkins, a mathematician. Both work in
California for a small company that develops motion detection for Wii
and PlayStation. The Australian team members are all based at the
University of Queensland’s Center for Advanced Imaging. Simone Bosshard
and Min Chen are neuroscientists who work with animal models of
epilepsy, while Quang Tieng is an applied mathematician.
The competition shows that sharing data to collaborate on complex
problems can yield clever and unexpected solutions. Also noteworthy is
that none of the winners were clinicians. This suggests that if we can
find a platform to engage players from other seemingly distant
disciplines, the chances are greater to find innovative, out-of-the-box
solutions to current challenges.
That’s why NINDS has launched iEEG.org to catalyze collaboration and
sharing of datasets, algorithms, and research tools for BigData studies
of epilepsy. On this website are almost 2,000 datasets freely available
for analysis, and there is already a user base of more than 670 active
collaborators. It will be exciting to see how long it takes to go even
beyond that 82 percent predictability.
Links:
National Institute of Neurological Disorders and Stroke: Epilepsy
IEEG.ORG, the NINDS-supported research initiative to catalyze collaboration and and data sharing for BigData studies of epilepsy
Litt Lab, University of Pennsylvania, Philadelphia