I was diagnosed with epilepsy in my teens. The daily barbiturate I was prescribed to prevent seizures made me sleepy and I had trouble concentrating. Plus, I suffered even worse headaches than the ones that sent me to the neurologist in the first place.
Thankfully, the diagnosis turned out to be wrong, and I ditched the hated meds. But the 50 million people worldwide living with epilepsy aren’t so lucky.
Most don’t know when a debilitating seizure will strike. As with my experience, the daily doses of medicines they take can have unpleasant or even dangerous side effects, with serious consequences for quality of life.
For example, those with epilepsy are often barred from driving. Many countries and all U.S. states require patients to prove they are seizure-free for a certain period of time before they can get behind the wheel.
“Our inability to predict when people will have seizures really limits how epilepsy patients can live their lives,” said Dr. Christian Meisel, a neurologist and researcher at the University of Dresden, in Germany.
Meisel thinks he can change that with GPUs and deep learning.
Warning System for Seizures
Meisel wants to create a seizure warning system for those with epilepsy, which could open up new possibilities for treating the disease.
For example, it may be possible for patients to take a drug only when they’re at risk of a seizure, Meisel said. Or there may be a way to electrically stimulate the brain to prevent a seizure. Some 30 percent of people with epilepsy don’t respond to drugs.
“If we could tell patients, ‘in the next few hours you have a higher chance of a seizure,’ it would help people plan their lives,” he said.
To forecast seizures, Meisel combines his theories about how seizures happen with deep learning models that detect changes in electrical activity in the brain that signal a seizure. Brain activity data comes from patients wearing an EEG device in which electrodes are placed directly on the surface of the brain.
He first tried this approach by entering a seizure-prediction contest on Kaggle, the data science competition platform, while he was a postdoctoral fellow at the U.S. National Institutes of Health. He continued the work in his current post, using NVIDIA TITAN X GPUs for training and inference.
In both experiments, his algorithm accurately forecasted seizures with a score of 0.8 – better than flipping a coin, which scores 0.5 – but not quite a perfect prediction of 1.
“What this tells you is that anticipating or predicting seizures is possible,” Meisel said. “That’s in stark contrast to what people thought.”
The Big Switch
Epilepsy is just the first step in Meisel’s larger goal of understanding the brain’s often rapid transition between a healthy state and a diseased state.
“Many diseases — from neuropsychiatric (depression, schizophrenia) to neurologic (multiple sclerosis) to autoimmune, plus other conditions like sepsis — are characterized by rapid escalation of symptoms,” Meisel said.
Today, doctors have two options: They can prescribe drugs that patients with these and similar conditions take continuously or they can treat the disease only when it’s acute.
“Neither of these is ideal,” Meisel said. “We need to do better than this.”